![4 1 Introduction to mathematical epidemiology
1.1 A brief history of infectious diseases
In the year 2020, everybody–at home on their couch–became an infectious disease
expert. To calibrate all this newly gained knowledge, it seems a good idea to briefly
reiterate the basic concepts and nomenclature of infectious diseases. Infectious dis-
eases are spread by either bacterial or viral agents and are ever-present in society
[8]. Every once in a while, this may result in outbreaks that have a significant impact
on a local or global level. Depending on their spatial and temporal spread, we can
classify outbreaks as endemic, epidemic, or pandemic [28]. Endemic outbreaks, for
example chickenpox, are permanently present in a region or a population. Epidemic
outbreaks, for example the seasonal flu, affects a lot of people in a short period of
time, spread across several communities, and then disappear. Pandemic outbreaks,
for example the Spanish flu, are epidemic outbreaks that affects a lot of people in a
short period of time and spread across the entire world. Mathematical modeling of
infectious diseases is important to understand their outbreak dynamics and inform
political decision making to manage their spread [4].
Table 1.1 History of recent infectious disease outbreaks. Time period, type of disease, number
of deaths, and location.
period disease deaths location
1346 - 1350 Black Death 100,000,000 1/3 of Europe
1665 - 1666 Great Plague 100,000 1/4 of London
1918 - 1920 Spanish flu 50,000,000 worldwide
1980 measles 2,600,000 worldwide / year
2003 SARS 774 worldwide
2009 - 2010 H1N1 18,500 worldwide
2011 tuberculosis 1,400,000 worldwide / year
HIV/AIDS 1,200,000 worldwide / year
malaria 627,000 worldwide / year
2011 measles 160,000 worldwide (-94%)
2012 - 2020 MERS 866 worldwide
seasonal flu 35,000 United States / year
2014 - 2016 ebola 11,000 Africa
2018 ebola 2,280 Congo
2019 -2021 COVID-19 3,300,000 worldwide
The COVID-19 pandemic. On March 11, 2020, Tedros Adhanom Ghebreyesus, the
Director-General of the World Health Organization, declared the COVID-19 outbreak
a global pandemic [49]. On that day, it had affected 126,702 people worldwide. What
nobody could have foreseen is that within the following year, by March 11, 2021,
this number had increased by three orders of magnitude, to 118.57 million [12].
Naturally, we often think of the COVID-19 pandemic as the deadliest and most
devastating infectious disease in modern history.](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-21-320.jpg)
![6 1 Introduction to mathematical epidemiology
Table 1.1 summarizes recent infectious disease outbreaks ranging from the Black
Death in the 14th century with an estimated 100 million deaths across Europe to the
current COVID-19 pandemic with 3.3 million deaths worldwide to date. Notably, in
1980, the annual death toll of the measles with 2.6 million was of comparable size.
Today, after massive vaccination campaigns, this number has dropped significantly,
by 94% to 160,000 [5].
Fig. 1.1 The Lessons of the Pandemic. In this famous Science publication, George A. Soper
reflects on the Spanish flu that resulted in 50 million deaths worldwide in 1918 and 1919.
Figure 1.1 shows the cover page of the Science publication The Lessons from the
Pandemic from May 1919, a reflection on the scientific understanding of the Spanish
flu [29]. While some of it is specific to the nature of the 1918 influenza and the time
during which it occurred, much of it still applies to the COVID-19 pandemic today:
The Spanish flu lasted from 1918 to 1920 and, similar to COVID-19, occurred in
multiple waves. Both, the Spanish flu and COVID-19, are contagious respiratory
illnesses that spread from person to person, mainly by droplets, through cough,
sneeze, or talk. Naturally, increased hygiene, mask wearing, and physical distancing
in addition to strict isolation and quarantine are successful strategies to manage both
conditions [10]. While COVID-19 is caused by a coronavirus, SARS-CoV-2, the
1918 influenza was caused by the H1N1 influenza A virus, a virus of avian origin.
Within 25 months, it infected 500 million people, one third of the world’s population,
and resulted in more than 50 million deaths [5]. The major differences between the
Spanish flu and COVID-19 are their high risk populations and the mechanisms of
death: In contrast to the Spanish flu, which affected mainly healthy adults between
25 and 40 years of age, COVID-19 affects mainly individuals of 65 years and older
with comorbidities. Victims of the 1918 influenza mainly died from secondary
bacterial pneumonia, whereas victims of COVID-19 die from an overactive immune
response that results in organ failure [32]. Nevertheless, comparing the COVID-
19 outbreak to previous pandemics can provide insight and guidance to manage
the current COVID-19 pandemic. An important element of this comparison are
mathematical and computational tools that have been designed to quantify and
explain the spreading mechanisms of infectious diseases. This is the objective of
mathematical and computational epidemiology [3].](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-23-320.jpg)
![1.2 Introduction to epidemiology 7
1.2 Introduction to epidemiology
Epidemiology is the study of distributions, patterns, and determinants of health-
related events in human populations [3]. It is a cornerstone of public health , and
shapes policy decisions by identifying risk factors for outbreaks and targets for
prevention [28].
Epidemiology literally means the study of what is upon the people. It is
derived from the greek words epi, meaning upon demos, meaning people, and
logos meaning study, suggesting that it applies only to human populations. By
this definition, epidemiology is the scientific, systematic, data-driven study of
the distribution, i.e., where, who, and when, and determinants, i.e., causes and
risk factors, of health-related patterns and events in specified populations, i.e.,
the world, a country, state, county, city, school, neighborhood. Epidemiology
also includes the study of outbreak dynamics, outbreak control, and informing
political decision making.
Agent, host, and environment. A key premise in epidemiology is that health-related
events are not evenly distributed in a population; rather, they affect some individuals
more than others [3]. An important goal of epidemiology is to identify the causes that
put these individuals at a higher risk [2]. A simple but popular model to analyze and
explain disease causation is the epidemiologic triangle. The epidemiologic triangle
summarizes the interplay of the three components that contribute to the spread of an
infectious disease: an external agent, a susceptible host, and an environment in which
agent and host interact [9]. Descriptive epidemiologists characterize this interaction
as the seed, the soil, and the climate [42]. Effective public health measures assess
all three components and their interactions to control or prevent the spreading of a
disease. Figure 1.2 illustrates the epidemiologic triangle of COVID-19 with agent,
host, and environment. For COVID-19, the agent is the SARS-CoV-2 virus, the
hosts are people, and the environment are droplets. Interventions between any two
of these three components can help reduce the spread of the disease [20]. For exam-
ple, reduced exposure, vaccination, and antiviral treatment can modulate agent-host
interactions and reduce the number of new infections.
From descriptive to mathematical epidemiology. Mathematical models of infec-
tious diseases date back to Daniel Bernoulli’s model for smallpox in 1760 [5], and
they have been developed and improved extensively since the 1920s [14]. In the
middle of the 19th century, the English physician John Snow conducted a famous
series of experiments of the cholera outbreak in London to discover the cause of the
disease and to prevent its recurrence [46]. Because his research illustrates the classic
sequence–from descriptive epidemiology and hypothesis generation, to mathemat-
ical epidemiology and hypothesis testing–John Snow is considered the father of
modern epidemiology. Figure 1.4 summarizes the experiments of John Snow that
ended the cholera outbreak in London by removing the handle of a public water](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-24-320.jpg)
![8 1 Introduction to mathematical epidemiology
Fig. 1.2 Epidemiologic triangle of COVID-19 with agent, host, and environment. The epidemi-
ologic triangle illustrates the interplay of the three components that contribute to the spread of a
disease: an external agent, a susceptible host, and an environment in which agent and host interact.
For COVID-19, the agent is the SARS-CoV-2 virus, the hosts are people, and the environment are
droplets. Interventions between any of these three components can reduce the spread of the disease.
Fig. 1.3 Daniel Bernoulli is considered the most famous
mathematician of the Bernoulli family, although he actually
studied medicine. He was born on February 8, 1700 in
Groningen, the Netherlands, and is best known for his
applications of mathematics to mechanics. To prove the
efficacy of vaccination against smallpox, he proposed the
first compartment model in 1760 [5]. It only had two
compartments, but demonstrated the potential of modeling
to understand the mechanisms of transmission, predict future
spread, and control the outbreak through vaccination. Daniel
Bernoulli died on March 27, 1782 in Basel, Switzerland.
pump. Today the field of epidemiology is considered a quantitative discipline that
uses rigorous mathematical tools of probability and statistics to develop and test
hypotheses to understand and explain health-related events [3]. In simple terms,
today’s epidemiologists count, scale, and compare. They count the number of cases;
scale this count by a characteristic population to define fractions; and compare the
evolution of these fractions over time [9]. Mathematical epidemiology has advanced
significantly throughout the past decades [11], and models have become more so-
phisticated and complex [39]. This implies that in epidemiology today, is often no
longer possible to solve epidemiological models analytically and in closed form [3].
From mathematical to computational epidemiology. Computational epidemiol-
ogy is a multidisciplinary field that integrates mathematics, computer science, and
public health to better understand central questions in epidemiology such as the](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-25-320.jpg)
![1.2 Introduction to epidemiology 9
Fig. 1.4 The English physician John Snow is considered one
of the founders of modern epidemiology, in part because of
his research during the cholera outbreak in London in 1954
[46]. John Snow lived from March 15, 1813 to June 16, 1858.
He questioned the widely accepted paradigm that cholera
would spread by polluted bad air. He postulated that water
would be the cause of the cholera outbreak in Soho, London
in 1854. By talking to local residents, he identified the source
of the outbreak as the public water pump on Broad Street.
Although his chemical and microscopic observations were
inconclusive, his spreading patterns of the cholera outbreak
were convincing enough to persuade the local council to
disable a street pump on Broad Street by removing its handle.
There is a common belief that this action marked the ending
of the outbreak.
spread of diseases or the effectiveness of public health interventions [22]. Real-time
epidemiology is a rapidly developing area within computational epidemiology that
seeks to support policy makers–in real time–as an outbreak is unfolding. An impor-
tant application of real-time epidemiology is disease surveillance, the data-driven
collection, analysis, and interpretation of large volumes of disease data from a va-
riety of sources. This information can help to evaluate the effectiveness of control
and preventative health measures [12]. Central to computational epidemiology is
the accurate knowledge of people who are affected by an infectious disease at any
given point in time. During the COVID-19 pandemic, for the first time in history,
this information has been collected conscientiously, shared publicly, updated daily,
and made freely available in real time [13].
Sensitivity and specificity. A diagnostic tests, for example, the nasal swab
test, can be inaccurate in two ways: a false positive result erroneously labels
a healthy person as infected, resulting in unnecessary quarantine and contact
tracing; a false negative result oversees an infected person, resulting in the
risk of infecting others.
sensitivity = true positives / all sick people
Sensitivity measures the proportion of positives, diseased people, that are
correctly identified and not overlooked.
specificity = true negatives / all healthy people
Specificity measures the proportion of negatives, healthy people, that are
correctly identified and not classified as diseased. A perfect test has 100%
sensitivity and 100% specificity.](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-26-320.jpg)
![1.4 The basic reproduction number 11
Fig. 1.5 Testing for COVID-19. The objective of COVID-19
testing is to probe whether an individual is currently or
has previously been infected with SARS-CoV-2. A nasal or
throat swab test is a diagnostic test that provides information
about an acute infection. Diagnostic testing is critical to
provide early treatment, quarantine individuals, and trace
and isolate their contacts to reduce the spread of the virus. A
blood test is an antibody test that provides information about
a previous infection. Antibody tests are critical to estimate
the overall dimension of an outbreak. Throughout this book,
we use COVID-19 case data from public dashboards based
on confirmed positive diagnostic tests.
An antibody test can detect antibodies that are made by your immune system in
response to a previous infection with SARS-CoV-2. In contrast to genetic material
from the virus or specific proteins made by the virus, antibodies can take several
days or even weeks to develop, but they can remain present for several weeks or
months after recovery. A serology test is an antibody test that looks for antibodies in
blood samples. These venous blood samples are typically collected at the doctor’s
office or in the clinic. Antibody tests should not be used to diagnose an active infec-
tion. Instead, antibody testing is important to understand the kinetics of the immune
response to infection, clarify whether an infection protects from future infection,
characterize how long immunity will last, and estimate the overall dimension of an
outbreak.
Table 1.2 contrasts the three most common types of tests for COVID-19. Throughout
this book, we use COVID-19 case data from public dashboards based on confirmed
positive diagnostic tests. Only in Chapters 8 and 13, where we explore the effects of
asymptomatic transmission, we use specialized models that combine case data from
diagnostic tests with seroprevalence data from antibody tests.
1.4 The basic reproduction number
The basic reproduction number is a powerful but simple concept to explain the
contagiousness and transmissibility of an infectious disease [6]. For decades, epi-
demiologists have successfully used the basic reproduction number to quantify how
many new infections a single infectious individual creates in an otherwise com-
pletely susceptible population [6]. During the COVID-19 pandemic, the public me-
dia, scientists, and political decision makers across the globe have adopted the basic
reproduction number as an illustrative metric to explain and justify the need for dif-
ferent outbreak control strategies [10]: An outbreak will continue for reproduction
numbers larger than one, R0 > 1, and come to an end for reproduction numbers
smaller than one, R0 < 1 [8]. However, especially in the midst of a global pandemic,
it is difficult–if not impossible–to measure R0 directly [48].](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-28-320.jpg)
![12 1 Introduction to mathematical epidemiology
Table 1.3 Basic reproduction numbers and herd immunity thresholds for common infectious
diseases. Herd immunity is the indirect protection from an infectious disease that occurs when a
large fraction of the population has become immune. The herd immunity threshold, H = 1 − 1/R0,
beyond which this protection occurs is a function of the basic reproduction number R0.
disease R0 H disease R0 H
measles 12 - 18 92 - 95% mumps 4.0 - 7.0 75 - 86%
pertussis 12 - 17 92 - 94% COVID-19 2.0 - 6.0 50 - 83%
rubella 6 - 7 83 - 86% SARS 2.0 - 5.0 50 - 80%
smallpox 6 - 7 83 - 86% ebola 1.5 - 2.5 33 - 60%
polio 5 - 7 80 - 86% influenza 1.5 - 1.8 33 - 44%
Table 1.3 summarizes the basic reproduction numbers for common infectious dis-
eases. It varies from R0 = 1.5 − 1.8, for less contagious diseases like influenza, to
R0 = 12 − 18 for the measles and pertussis. Knowing the precise value of R0 is
important, but challenging, because of limited testing, inconsistent reporting, and
incomplete data [6]. Throughout this book, instead of measuring the basic repro-
duction number directly, we estimate it using mathematical modeling and reported
case data [17]. Mathematical models interpret the basic reproduction number R0
as the ratio between the infectious period C, the period during which an infectious
individual can infect others, and the contact period B, the average time it takes to
come into contact with another individual [8],
R0 = C/B . (1.1)
The longer a person is infectious, and the more contacts the person has during this
time, the larger the reproduction number [6]. While we cannot control the infectious
period C, we can change our behavior to increase the contact period B [17]. This is
precisely what community mitigation strategies and political interventions seek to
attempt.
The reproduction number of COVID-19. Since the beginning of the coronavirus
pandemic, no other number has been discussed more controversially than the re-
production number of COVID-19 [22]. The earliest COVID-19 study that followed
the first 425 cases of the Wuhan outbreak via direct contact tracing reported a basic
reproduction number of R0 = 2.2 [15]. However, especially during the early stages
of the outbreak, information was limited because of insufficient testing, changes
in case definitions, and overwhelmed healthcare systems. While the concept of R0
seems fairly simple, the reported basic reproduction numbers for COVID-19 vary
hugely with country, culture, calculation, and time [17]. Most basic reproduction
numbers of COVID-19 we see in the public media today are estimates of mathemat-
ical models. These estimates depend critically on the choice of the model, its initial
conditions, and many other modeling assumptions [6]. To no surprise, the mathe-
matically predicted basic reproduction numbers cover a wide range, from R0 = 2−4
for exponential growth models to R0 = 4 − 7 for more sophisticated compartment
models [22].](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-29-320.jpg)
![1.4 The basic reproduction number 13
Throughout this book, we identify or infer the reproduction number of COVID-
19–and the uncertainty associated with it–using computational epidemiology [22],
Bayesian analysis [10], and reported case data [12]: In the first example for the very
early COVID-19 outbreak in China in Section 7.8, we identify a basic reproduction
number of R0 = 12.58±3.17 across 30 Chinese provinces. In the second example for
the early outbreak in the United States in Section 10.2, we find a basic reproduction
number of R0 = 5.30 ± 0.95 across all 54 states and territories. This value suggests
that COVID-19 is less contagious than the measles and pertussis with R0 = 12 − 18,
as infectious as rubella, smallpox, polio, and mumps with R0 = 4 − 7, slightly more
infectious than SARS with R0 = 2 − 5, and more infectious than an influenza with
R0 = 1.5 − 1.8. Our basic reproduction number for the United States is significantly
lower than our basic reproduction number for China, which could be caused by
an increased awareness of COVID-19 transmission a few weeks into the global
pandemic. In our third example for the early outbreak in Europe in Section 12.2, the
basic reproduction number takes similar values of R0 = 4.62 ± 1.32 across all 27
countries.
During these early stages of exponential growth, with new case numbers doubling
within two or three days, the most urgent question amongst health care providers
and political decision makers was: Can we reduce the reproduction number? For the
broad public, this question became famously and illustratively rephrased as: Can we
flatten the curve [15]? For the modeling community, the quest for a lower reproduc-
tion number all of a sudden meant that traditional epidemiology models were no
longer suitable because of changes in the disease dynamics [12]. While traditional
models with static parameters were well-suited to model the outbreak dynamics
of unconstrained, freely evolving infectious diseases with fixed basic reproduction
numbers in the early 20th century, they fail capture how behavioral changes and po-
litical interventions can modulate the reproduction number to manage the COVID-19
pandemic in the 21st century [17]. In fact, static reproduction numbers are probably
the single most common cause of model failure in COVID-19 modeling [12].
Fortunately, several months into the pandemic, most countries have successfully
managed to flatten the new-case-number curves and the reproduction numbers have
dropped to values closer to or below one. To model these changes in disease dynamics
and reproduction, in Section 7.7, we introduce a dynamic reproduction number,
R(t), that accounts for time-varying contact periods. In our fourth example for the
European Union in Section 12.2, we show that the initial basic reproduction number
of R0 = 4.22±1.69 dropped to an effective reproduction number of R(t) = 0.67±0.18
by mid May 2020. Using machine learning, we correlate mobility and reproduction
and identify the responsiveness between the drop in air traffic, driving, walking,
and transit mobility and the drop in reproduction to ∆t = 17.24 ± 2.00 days. In
the final examples in Sections 13.2 and 13.3 of nine locations across the world,
we systematically infer the dynamic reproduction number R(t) throughout a time
window of 100 days while accounting for both symptomatic and asymptomatic
transmission.
From the failure of traditional static epidemiology models [15], we have now
learned that we need to introduce dynamic time-varying model parameters if we](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-30-320.jpg)
![14 1 Introduction to mathematical epidemiology
Fig. 1.6 Basic reproduction numbers and herd immunity thresholds for common infectious
diseases. Herd immunity is the indirect protection from an infectious disease that occurs when a
large fraction of the population has become immune, either through previous infection or through
vaccination. The black line highlights the herd immunity threshold, H = 1 − 1/R0, beyond which
this protection occurs, as a function of the basic reproduction number R0. The herd immunity
threshold varies from 33-44% for influenza to 92-95% for the measles.
want to correctly model behavioral and political changes and reproduce the reported
case numbers [17]. This naturally introduces a lot of freedom, a large number of
unknowns, and a high level of uncertainty. However, in stark contrast to the epidemic
outbreaks in the early 20th century, we now have thoroughly-reported case data and
the appropriate tools [23] to address this challenge. The massive amount of COVID-
19 case data, well documented and freely available, has induced a clear paradigm
shift from traditional mathematical epidemiology towards data-driven, physics-based
modeling of infectious disease [1]. This new technology naturally learns the most
probable model parameters–in real time–from the continuously emerging case data,
allows us to make projections into the future, and quantifies the uncertainty on the
estimated parameters and predictions [26].
1.5 Concept of herd immunity
An important consequence of the basic reproduction number R0 is the condition
for herd immunity [9]. Herd immunity describes the indirect protection from an
infectious disease that occurs when a large fraction of the population has become
immune, either through previous infection or through vaccination [3]. The critical
threshold at which the disease reaches this endemic steady state is called the herd
immunity threshold,
H = 1 − 1/R0 . (1.2)
The larger the basic reproduction number R0, the higher the herd immunity threshold
H. Table 1.3 and Figure 1.6 summarize the basic reproduction numbers R0 for](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-31-320.jpg)
![1.6 Concept of immunization 15
several common infectious diseases along with the estimates of their herd immunity
thresholds H. The herd immunity threshold varies from 33-44% for influenza to 92-
95% for the measles. For the reported basic reproduction numbers of R0 = 2.0 − 6.0
of COVID-19, the estimated herd immunity threshold would range from 50-83%.
Recent studies that account for the emerging new and more infectious B.1.351 and
B.1.1.7 variants of COVID-19 estimate these values to 75-95% [20].
1.6 Concept of immunization
To prevent or revert an epidemic outbreak, we need to ensure that, on average,
every infectious individual infects less than one new individual. The concept of herd
immunity describes the natural path towards ending an outbreak. However, if the
basic reproduction number R0 is large, the herd immunity threshold H = 1 − 1/R0 is
high, and waiting for herd immunity through infection alone can be quite devastating.
Vaccination is a powerful strategy to accelerate the path towards herd immunity [1].
Fig. 1.7 Vaccination against COVID-19. The objective of
COVID-19 vaccination is to provide immunity against severe
acute respiratory syndrome coronavirus 2 or SARS-CoV-2,
the virus that causes COVID-19 [30]. By May, 2021,
thirteen vaccines were authorized for public use: two RNA
vaccines (Pfizer–BioNTech and Moderna), five conventional
inactivated vaccines (BBIBP-CorV, CoronaVac, Covaxin,
WIBP-CorV and CoviVac), four viral vector vaccines
(Sputnik V, Oxford–AstraZeneca, Convidecia, and Johnson
& Johnson), and two protein subunit vaccines (EpiVacCorona
and RBD-Dimer); and more than a billion doses of COVID-
19 vaccines were administered worldwide.
Immunization threshold. Vaccination effectively reduces the susceptible population
S. Successfully immunizing a fraction I of the population, reduces the susceptible
population from S to S∗
0 = [ 1 − I ] S0 and, with it, the reproduction number from R0
to R∗
0 = [ 1 − I ] R0. The critical immunization threshold I, below which the effective
basic reproduction number is smaller than one, R∗
0 = [ 1 − I ] R0 < 1, defines the
fraction of the population that needs to be immunized to prevent or revert the outbreak
of an epidemic,
I > 1 − 1/R0 . (1.3)
From Table 1.3, we conclude that the required immunization fraction varies signif-
icantly, from I > 92 − 95% for the measles with a basic reproduction number of
R0 = 12 − 18 to I > 33 − 44% for the common influenza with a basic reproduction
number of R0 = 1.5−1.8. This explains, at least in part, why some infectious diseases
are a lot more difficult to control through immunization than others.](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-32-320.jpg)
![16 1 Introduction to mathematical epidemiology
Eradication through vaccination. Once enough individuals are immunized–either
through infection or through vaccination–the outbreak stops. A disease that stops
circulating in a specific region is considered eliminated in that region. Polio, for
example, was eliminated in the United States by 1979 after widespread vaccination
efforts. A disease that is eliminated worldwide is considered eradicated. Eradicating
a disease through vaccination is a desirable but elusive goal [1]. Malaria has been
a candidate for eradication, but although its incidence has been drastically reduced
through vaccination, completely eradicating it remains challenging because infection
does not result in life long immunity. Polio has been eliminated in most countries
through massive vaccination efforts, but still remains present in some regions be-
cause its early symptoms often remain unnoticed and infected individuals continue to
infect others. Measles have been the target of widespread vaccination, but although
the disease is highly recognizable through its characteristic rash, a long latent pe-
riod from exposure to the first onset of symptoms complicates outbreak control. To
this day, smallpox is the only human infectious disease that has been successfully
eradicated through vaccination. The eradication of smallpox is the result of focused
surveillance, rapid identification, and ring vaccination [8]. In a massive vaccina-
tion campaign launched in 1967, anyone who could have possibly been exposed to
smallpox was quickly identified and vaccinated to prevent its further spread. The
last known case of smallpox occurred in Somalia in 1977. In 1980 World Health
Organization declared smallpox eradicated. The eradication of smallpox remains one
of the most notable and profound public health successes in history.
Efficacy and risk ratio. The efficacy e of a vaccine is the relative reduction in the
disease attack rate between the unvaccinated placebo group npla and the vaccinated
group nvac. For a randomized trial with an equal allocation, meaning equally sized
placebo and vaccinated groups, npla = nvac, the efficacy is
e =
npla − nvac
npla
· 100% =
1 −
nvac
npla
· 100% = 1 − r . (1.4)
The risk ratio r is the ratio between the attack rate of the vaccinated group nvac and
the placebo group npla,
r =
nvac
npla
· 100% = 1 − e . (1.5)
The efficacy is an important measure to characterize the success of a vaccine and
define critical thresholds below which a vaccination trial should stop.
Table 1.4 Contingency table to quantify the significance of a vaccine. The table compares the
total number of vaccinated and placebo individuals that developed and did not develop the disease.
positive negative total
vaccine a b a+b
placebo c d c+d
total a+c b+d n](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-33-320.jpg)
![1.6 Concept of immunization 17
Example: Efficacy of the first COVID-19 vaccine. On November 9, 2020, the
Pfizer and BioNTech trial reported a number or COVID-19 cases of ncovid =
npla + nvac = 94 and an efficacy of e emin with emin = 90%. The efficacy
e of a vaccine is the relative reduction in the disease attack rate between the
unvaccinated placebo group npla and the vaccinated group nvac,
e =
1 −
nvac
npla
· 100% = 1 − r with r =
nvac
npla
· 100% = 1 − e .
where r is the risk ratio. The Pfizer and BioNTech trial was a randomized trial
with an equal allocation [30]. Although it did not report detailed numbers, we
can estimate the number of placebo and vaccinated cases npla and nvac and the
risk ratio r from the reported efficacy emin = 90% with nvac = ncovid − npla,
e =
1 −
ncovid − npla
npla
· 100% =
2 −
ncovid
npla
· 100% emin .
Solving for the number of placebo cases yields the general equation,
npla
ncovid
2 − emin
,
and, for the Pfizer and BioNTech case, npla ncovid/1.1 = 85.45. This implies
that for a total of ncovid = 94 COVID-19 positive cases, at a number of placebo
cases npla = 86 and vaccinated cases nvac = 8, the efficacy is larger than 90%.
Back-calculating the efficacy for these populations,
e =
1 −
nvac
npla
· 100% =
1 −
8
86
· 100% = 90.7% 90% = emin ,
confirms the simulation. According to the protocol, Pfizer and BioNTech
planned to take a look at the data at five stages with ncovid = 32,64,92,120,164
reported positive cases and only continue the trial if the efficacy was above
e ecrit with ecrit = 62.7%. From this information, we can calculate the
critical numbers npla and nvac below which the trial would have stopped at any
of the five stages. Solving for the number of placebo cases yields the general
equation,
npla
ncovid
2 − ecrit
,
and, for the Pfizer and BioNTech case, npla ncovid/1.373. This implies that for
ncovid = 32,64,92,120,164, the minimum number of placebo cases to continue
the trial was npla ≥ 24,47,68,88,120, for which the resulting efficacies of
e = 66.7%,63.8%,64.7%,63.6%,63.3% would all have been slightly above
the critical threshold of ecrit = 62.7%.](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-34-320.jpg)
![1.7 Mathematical modeling in epidemiology 19
positive negative total
vaccine 8 21,761 21,769
placebo 86 21,683 21,769
total 94 43,444 43,538
We can use the contingency table to estimate the efficiency of the Pfizer BioN-
Tech vaccine. Fisher’s exact test calculates the significance of the deviation
from a null hypothesis,
p =
21769! 21769! 94! 43444!
8! 21761! 86! 21683! 43538!
,
and confirms, for the Pfizer and BioNTech case with p 0.00001, that indi-
viduals in the vaccinated and placebo groups will not equally likely contract
COVID-19.
1.7 Mathematical modeling in epidemiology
During the early onset of the COVID-19 pandemic, all eyes were on mathematical
modeling with the general expectation that mathematical models could precisely
predict the trajectory of the pandemic. Mathematical modeling rapidly became front
and center to understanding the exponential increase of infections, the shortage of
ventilators, and the limited capacity of hospital beds; too rapidly as we now know.
Bold and catastrophic predictions not only initiated a massive press coverage, but
also a broad anxiety in the general population [15]. However, within only a few
weeks, the vastly different predictions and conflicting conclusions began to create
the impression that all mathematical models are generally unreliable and inherently
wrong [12]. While the failure of COVID-19 modeling–often by an order of mag-
nitude and more–was devastating for policymakers and public health practitioners,
initial mistakes are not new to the modeling community where an iterative cycle of
prediction, failure, and redesign is common standard and best practice [26]. However,
the successful use of mathematical models implies to set the expectations right [45].
Understanding what models can and cannot predict is critical to the Art of Mod-
eling. Epidemiologists distinguish two kinds of models to understand the outbreak
dynamics of an infectious disease: statistical models and mechanistic models [12].
Depending on the degree of complexity, the most popular mechanistic models are
compartment models and agent-based models.
Statistical models, or more precisely, purely statistical models, use machine learn-
ing or regression to analyze massive amounts of data and project the number of
infections into the future. The essential idea is to select a function D(t), use statisti-
cal tools to fit its coefficients to reported case data D̂(t), and make projections into
the future. The function can be quadratic, cubic, logistic, power-law, or exponential,](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-36-320.jpg)
![20 1 Introduction to mathematical epidemiology
Table 1.5 Mathematical modeling of COVID-19. Summary of the three most common models.
Statistical or forecasting models fit nonlinear functions to case data over time, whereas mechanistic
models, including compartment and agent-based models, simulate outbreak and contact dynamics.
Throughout this book, we use compartment models to simulate the outbreak dynamics of infectious
diseases including COVID-19.
statistical models mechanistic models
forecasting models compartment models agent-based models
use machine learning, statis-
tics, regression, or method of
least squares
use physics-based modeling
based on nonlinear reaction-
diffusion equations
use rule-based approaches to
study the interaction of au-
tonomous systems
model case numbers through
a nonlinear function
model population through
compartments
model every individual as an
independent agent
formulate number of cases as
a function of time
formulate rules by which in-
dividuals pass through the
compartments
formulate simple rules by
which individual agents in-
teract
fit coefficients that are purely
phenomenological
infer parameters that have a
mechanistic interpretation
identify parameters that
summarize human behavior
predicts case numbers from
fitting a function
predict outbreak dynam-
ics, characterize sensitivi-
ties, quantify uncertainties
predict outbreak dynamics
as emergent collective be-
havior of individual agents
no feedback mechanisms nonlinear feedback discrete contact networks
predictions are inexpensive,
but very unreliable
predictions are reliable only
for a small time window
predictions are detailed, but
computationally expensive
for example, D(ϑ,t) = exp(c0 + c1 t + c2 t2 + c3 t3), where D(ϑ,t) are the modeled
cumulative cases per day, ϑ = {c0,c1,c2,c3} are the model parameters, and t is
the time. One of the simplest statistical tools to compare the model D(ϑ,t) to the
data D̂(t) and identify values for its parameters ϑ is the method of least squares.
It is important to understand that these parameters are purely phenomenological,
they are derived purely by fitting a curve, and typically do not have a mechanistic
interpretation. Early in an outbreak, when little is known about disease transmission,
epidemiologists often use statistical models because they do not rely on any prior
knowledge of the disease. An example for a statistical model of COVID-19 is the ini-
tial IHME model [23]. Early in the pandemic, the IHME model used case data from
China and Italy to create similar curves, forecast case numbers in the United States,
and inform the White House’s response to the pandemic. Carefully constructed sta-
tistical frameworks can be used for short-time forecasting using machine learning or
regression. This could potentially be useful to understand how to allocate resources
or make rapid short-term recommendations. However, purely statistical models can
neither capture the dynamics of disease transmission nor the effects of mitigation
strategies. This explains, at least in part, why the early COVID-19 predictions based
on purely statistical models were off by an order of magnitude or more. To address
these serious limitations, several COVID-19 models have now been adjusted to com-
bine both statistical modeling and mechanistic modeling.
Mechanistic models simulate the outbreak through interacting disease mechanisms
by using local nonlinear population dynamics and global mixing of populations](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-37-320.jpg)
![1.7 Mathematical modeling in epidemiology 21
[4]. The underlying idea is to identify fundamental mechanisms that drive disease
dynamics, for example the duration of the infectious period or the number of con-
tacts an infectious individual has during this time. Unlike purely statistical models,
mechanistic models include important nonlinear feedback: The more people be-
come infected, the faster the disease spreads. By their very nature, the parameters
of mechanistic models are not just fitting parameters, they usually have a clear epi-
demiological interpretation. This makes mechanistic modeling a powerful strategy
to explore different outbreak scenarios or study how an outbreak would change
under various assumptions and political interventions [12]. Another advantage of
mechanistic models is that we can adjust and improve them dynamically as more
information becomes available. Throughout this book, we gradually improve a class
of mechanistic models by adding new information. For example, we introduce time-
varying dynamic contact rates that vary in different lockdown levels and add the
effect of asymptomatic transmission [26]. Even if we do not precisely know the
dimension of asymptomatic disease spread, we can use mechanistic models to study
what-if scenarios: What would the disease landscape look like if two third of all
infectious were asymptomatic? Mechanistic modeling naturally extends into sensi-
tivity analysis and uncertainty quantification. As such, it not only provides valuable
information about the robustness of the model, but also about the most effective
parameters to modulate a disease outbreak [19]. Rather than studying a single one
disease trajectory, we could explore a range of trajectories around the mean and
characterize the best- and worst-case scenarios. The two most popular mechanistic
models in epidemiology are compartment models and agent-based models, and both
have been used to understand the outbreak dynamics of COVID-19. When choosing
between compartment models and agent-based, it is important to understand the
major strengths and weaknesses of each model.
Compartment models are the most common approach to model the epidemiology
of an infectious disease [14]. Compartment models simulate the collective behavior
of subgroups of the population through a number of compartments with labels, for
example, SEIR for susceptible, exposed, infectious, and recovered. Individuals move
between compartments and the order of the labels indicates the successive motion,
for example, SEIS means susceptible, exposed, infectious, then susceptible again [4].
The underlying principle is to model the time evolution of these groups through a set
of coupled ordinary differential equations, identify rate constants that characterize
their interaction using reported case data, and vary these rate constants to probe
different outbreak scenarios.
Figure 1.8 illustrates a compartment model that represents the characteristic time-
line of COVID-19 through six compartments, the susceptible, exposed, infectious,
recovered, hospitalized, and dead groups. This SEIRHD model is defined through a
set of six ordinary differential equations that simulate how many individuals reside
in each compartment throughout the duration of the outbreak. The model parameters
of a compartment model define the transition rates between the individual compart-
ments and, for more complex models, the fraction of individuals that transition into
a particular path of the disease. For this example, the parameters ϑ = {β,α,γ,νh,νd}](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-38-320.jpg)
![22 1 Introduction to mathematical epidemiology
Fig. 1.8 Characteristic timeline of COVID-19. On day 0, a fraction of all susceptible individuals
is exposed to the virus. After a latent period of A = 1/α = 3 days, the exposed individuals become
infectious. After an infectious period of C = 1/γ = 10 days, a fraction (1 − νh) transition to
the recovered group, whereas a fraction νh develops severe symptoms and is hospitalized. Of the
hospitalized individuals, a fraction (1 − νd) recovers, whereas a fraction νd becomes dead. We can
simulate this behavior through an SEIRHD model with six compartments.
are the contact rate β, latent rate α, and infectious rate γ, and the hospitalized and
dead fractions, νh and νh · νd. From reported case numbers, hospitalizations, and
deaths, we can identify or infer the set of model parameters ϑ, the rate constants
and fractions, that best explain the model output using statistical tools. Importantly,
in contrast to purely statistical models, compartment models are based on model
parameters that have a clear physical interpretation. The most important parameters
of any compartment model are the contact rate β and the infectious rate γ. Together,
they define an important nonlinear feedback that is not present in purely statistical
models: The more people become infected, the faster the spread of the disease. The
basic reproduction number R0 = β/γ, the ratio of the contact rate β and the infec-
tious rate γ, characterizes the magnitude of this feedback. It is an easy-to-understand
disease metric that explains how quickly susceptible individuals become infected
and how fast a disease spreads across a population [6]. Some epidemiologists ar-
gue that, because of their mechanistic nature, compartment models are better suited
for long-term predictions than purely statistical models. While this might be true
for infectious diseases that develop freely, without any political intervention, the
COVID-19 pandemic has taught us that long-term predictions of outbreak dynamics
are challenging, even with the most sophisticated mechanistic models [15]. Under-
standing the potential and the limitations of compartment modeling is one of the
main objectives of this book.
In Chapters 2 and 3, we introduce two simple classical compartment models be-
fore we introduce the most common compartment model for COVID-19 in Chapter
4. We show how we can use these models to estimate the reproduction number
from reported COVID-19 case data. Knowing the precise reproduction number has
important consequences for estimating the dimensions of herd immunity and im-
munization. Compartment models capture the fundamental dynamics of disease
transmission and the effects of public health interventions; however, classical com-
partment models ignore the dynamics of the contact rate and its variation across
a population. In Section 7.7 we introduce a compartment model that explicitly ac-
counts for a time-varying dynamic contact rate, β(t), and captures a varying contact
behavior in different subgroups of the population. The dynamic nature of the contact](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-39-320.jpg)
![1.8 Data-driven modeling in epidemiology 23
rate naturally introduces a dynamic reproduction number, R(t) = β(t)/γ and allows
us to quantify the effectiveness of policy measures as we discuss in Section 12.2.
Agent-based models simulate individuals or agents interacting in various social
settings and estimate the spread of a disease as these agents come into contact with
one another. The underlying idea is to represent each agent individually, formulate
relatively simple rules by which individual agents interact, and interpret the collec-
tive behavior across all agents as the emergent dynamics of an outbreak. A strength
of agent-based models is that they simulate human behavior very granularly: They
can assign different parameters or behavior patterns to each individual agent instead
of simulating the collective behavior of entire populations. As such, agent-based
models offer a lot more freedom than compartment models, but also require a lot
more detail. For example, to formulate rules of interaction, agent-based models draw
on social connectivity networks, from activity surveys, cell phone locations, public
transportation, or airlines statistics. By their very nature, agent-based approaches are
computationally expensive, especially for large populations. For small populations,
agent-based modeling is a powerful strategy to predict how individual behavior, for
example the violation of quarantine, leads to a collective behavior and modulates
disease spread. For large populations, agent-based modeling can help rationalize
collective model parameters, for example contact rates or reproduction numbers,
that feed into more abstract, population level models. Above a certain population
size, agent-based models simply become computationally unfeasible and most epi-
demiologists would turn to a more macroscopic approach that represents groups of
individuals collectively as subgroups of the population. Since our objective is to
design and discuss data-driven models for the COVID-19 pandemic, throughout the
remainder of this book, we focus exclusively on compartment models.
When choosing between purely statistical, compartment, and agent-based models,
it is important to know upfront which questions the model should address [45]. Ta-
ble 1.5 compares the three models and summarizes their strengths and weaknesses.
Throughout this book, we focus on mechanistic compartment modeling.
1.8 Data-driven modeling in epidemiology
One year after the World Health Organization had declared the COVID-19 outbreak a
global pandemic, SARS-CoV-2 has resulted in more than 118 million reported cases
across more than 180 countries and over 2.6 million deaths worldwide. Unlike any
other disease in history, the COVID-19 pandemic has generated an unprecedented
volume of data, well documented, continuously updated, and broadly available to
the general public. There is a critical need for time- and cost-efficient strategies
to analyze and interpret these data to systematically manage the pandemic on a
global level. Yet, the precise role of physics-based modeling and machine learning
in providing quantitative insight into the dynamics of COVID-19 remains a topic of
ongoing debate.](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-40-320.jpg)
![24 1 Introduction to mathematical epidemiology
Physics-based modeling is a successful strategy to integrate multiscale, multi-
physics data and uncover mechanisms that explain the dynamics of specific outbreak
characteristics. However, physics-based modeling alone often fails to efficiently
combine large data sets from different sources and different levels of resolution [26].
Machine learning is as a powerful technique to integrate multimodality, multifidelity
data, from cities, counties, states, and countries across the world, and reveal cor-
relations between different disease phenomena. However, machine learning alone
ignores the fundamental laws of physics and can result in ill-posed problems or non-
physical solutions [1]. Throughout this book, we illustrate how data-driven modeling
can integrate classical physics-based modeling and machine learning to infer critical
disease parameters–in real time–from reported case data to make informed predic-
tions and guide political decision making. As a valuable by product, this approach
naturally lends itself in sensitivity analysis and uncertainty quantification. From the
COVID-19 pandemic, we have learnt that even small inaccuracies in the model can
trigger large changes in the number of cases. To understand the vulnerability of the
model to these small changes, especially in view of the varying reporting practices
of the COVID-19 case data, sensitivity analysis and quantifying uncertainty have
become critical elements of robust predictive modeling.
Epidemiology data. In data-driven modeling, we need data to fit or infer our model
parameters. Unlike earlier pandemics for which case data are often sparse, irregular,
or incomplete, the COVID-19 pandemic is amazingly well documented [13]. On
hundreds of public COVID-19 dashboards, we can find and download a vast variety
of case numbers: daily new cases, active cases, recovered cases, seriously critical
cases, cumulative cases, and deaths, at the level of cities, counties, states, countries,
or the entire world [9, 12, 25]. Local dashboards often also share the number of hos-
pitalizations and intensive care units, which were of great concern especially at the
early onset of the pandemic. More recently, these dashboards have also included the
number of tests and vaccines. Seroprevalence data with information about the history
of the disease, both asymptomatic and symptomatic, are rare and often only avail-
able from scientific publications rather than governmental databases. These data are
typically updated on a daily basis, and contain notable weekday-weekend alterations.
When using the data to infer model parameters and learn about the outbreak behav-
ior, we usually smoothen these alteration using seven-day moving averages. Finally,
to compare data from different locations, we typically scale the reported case data
by the population. A common metric for comparison is the seven-day-per-100,000
incidence, the number of new cases per 100,000 individuals across a seven-day win-
dow [38]. Policy makers across the globe use this incidence value to characterize the
severity of the outbreak and justify the need for political interventions.
Figure 1.9 illustrates a typical data set for the COVID-19 outbreak that we use to
infer our epidemiological model parameters. The orange lines summarize the out-
break dynamics worldwide throughout the year after the World Health Organization
had declared COVID-19 a global pandemic, from March 11, 2020 to March 11,
2021. The light orange lines represent the reported daily new cases, which display
notable weekday-weekend fluctuations associated with testing and reporting irreg-](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-41-320.jpg)
![1.8 Data-driven modeling in epidemiology 25
Fig. 1.9 Daily new cases of COVID-19 and their seven-day moving average. Daily new COVID-
19 cases are reported on public dashboards worldwide. The light orange curve represents the raw
data, the dark orange curve is the seven-day moving average, both reported throughout the year
after the WHO had declared COVID-19 a global pandemic on March 11, 2020.
ularities. The dark orange lines are the associated seven-day moving average ˆ
I(t),
which smoothens the fluctuations and displays a much clearer trend in the disease
dynamics. Within this one-year window, the absolute daily case numbers peaked
on January 8, 2021 with 841,304 new cases. The moving seven-day average peaked
on January 11, 2021 with ∆Imax = 745,404 cases. If we assume an infectious pe-
riod of C = 1/γ = 7 days, this would result in a maximum infectious population
of Imax = 7 · ∆Imax = 7 · 745,404 = 5,217,828, meaning that mid January 2021,
more than 5 million people were sick with COVID-19. For a total population of
N = 7.8 billion people, this corresponds to a peak seven-day-per-100,000 incidence
of Imax · 100,000/N = 67. There are many different ways to use the reported case
data. The simplest way is to select a function, use statistical tools to fit its coefficients
to the reported case data, and make projections into the future. However, not only the
daily new cases, but also the seven-day average in Figure 1.9 display substantial fluc-
tuation and it seems difficult to find a function that could explain the orange curves.
For data-driven modeling, it is often easier to use the total cumulative case numbers,
which always increase monotonically and tend to be more smooth in general.
Figure 1.10 shows the total cumulative cases of COVID-19 D̂(t) and a simple
statistical model D(t) to fit the data throughout the first year of the pandemic.
During the very early stages of an outbreak, an exponential growth model, D(t) =
D0 exp(G t), with a growth rate G often provides a good approximation of the total
number of cases D̂(t) and is easy to fit. Indeed, this, or similar exponential models,
is what many early approaches used. While the initial phase of the outbreak is
well represented by exponential growth models, they soon tend to overestimate the
outbreak. This is why, during the early COVID-19 pandemic, there was a broad
overestimate of the number of cases and of the number of ventilators and hospital
beds needed to tread diseased individuals [15]. In this book, instead of using purely
statistical models, we use mechanistic models like the compartment model in Figure](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-42-320.jpg)
![26 1 Introduction to mathematical epidemiology
Fig. 1.10 Total cumulative cases of COVID-19 and a statistical model to fit the data. Cumulative
COVID-19 cases, the sum of all daily new cases to that date, are reported on public dashboards
worldwide. The solid orange curve represents the raw data, the dashed orange curve is an example
of a statistical model, both reported throughout the year after the WHO had declared COVID-19 a
global pandemic on March 11, 2020.
1.8. For this SEIRHD compartment model, the number of total cases is the sum
of all infectious, hospitalized, recovered and dead individuals, I(t) + R(t) + H(t) +
D(t). Throughout this book, we use different types of compartment models, infer
their model parameters using reported COVID-19 case data from public databases
[9, 12, 25], seroprevalence data from scientific publications [26], and mobility data
[2, 10, 10], and vary the parameters to probe different outbreak scenarios.
Problems
1.1 Testing, testing, testing. On June 6, 2020, Donald J. Trump, the President of
the United States, famously said “Remember this, when you test more, you have
more cases.” What is wrong with this statement? What did he really mean to say?
Discuss the implications of the testing frequency on reported case numbers, and,
ultimately, on policy making.
1.2 Herd immunity. On September 15, 2020, Donald J. Trump, the President of the
United States, claimed that COVID-19 would go away, without a vaccine. “You’ll
develop herd. Like a herd mentality.” What did he really mean to say? Discuss the
effects of vaccination on herd immunity.
1.3 Herd immunity. Assume the basic reproduction number of the swine flu, caused
by the H1N1 virus, was on the order of R0 = 1.4 − 1.6. Calculate its herd immunity
threshold H and compare it against the herd immunity thresholds of other infectious
diseases in Table 1.3 and Figure 1.6. Comment on whether this makes the swine flu
a good candidate for eradication through vaccination.](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-43-320.jpg)
![1.8 Data-driven modeling in epidemiology 27
1.4 COVID-19 variants B.1.351 and B.1.1.7. Assume the basic reproduction num-
ber of COVID-19 in December 2020 was R0 = 2.0. Calculate the herd immunity
threshold H. Now, assume that the variant B.1.351 with a 50% increased basic re-
production number has been introduced into the population [20]. How does the herd
immunity threshold change? How would the variant B.1.1.7 with a 56% increased
basic reproduction number change the herd immunity threshold? Comment on your
results.
Fig. 1.11 The Lessons of the Pandemic. Twelve condensed rules to the the avoidance of unneces-
sary personal risks and to the promotion of better personal health by George A. Soper in reflection
of the Spanish flu in 1919.
1.5 The Lessons of the Pandemic. In his Science publication The Lessons of the
Pandemic, George A. Soper lists twelve public health measures to manage the Spanish
flu in 1919. Read the twelve recommendations in Figure 1.11 and comment on which
of them are still valid for the COVID-19 pandemic today.
1.6 Efficacy of the AstraZeneca COVID-19 vaccine. The randomized, equal allo-
cation AstraZeneca trial enrolled 11636 participants in its interim efficacy analysis,
5807 in the vaccinated group and 5829 in the unvaccinated placebo group. Of the
vaccinated group, 30 developed COVID-19, and of the placebo group 101. Calculate
the efficacy and risk ratio of the AstraZeneca vaccine.
1.7 Case rates during the AstraZeneca COVID-19 trial. The AstraZeneca trial en-
rolled 11636 participants in its interim efficacy analysis, 5807 in the vaccinated
group and 5829 in the unvaccinated placebo group. Of the vaccinated group, 30
developed COVID-19, and of the placebo group 101. Create a contingency table for
the interim efficacy analysis. Calculate the overall case rate of the trial and the case
rates in the vaccinated and placebo groups.
1.8 Efficacy of the Janssen COVID-19 vaccine. The randomized, equal allocation
Janssen COVID-19 trial enrolled 39,321 participants, 19,630 received the vaccine](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-44-320.jpg)
![Chapter 2
The classical SIS model
Abstract The SIS model is the simplest compartment model with only two pop-
ulations, the susceptible and infectious groups S and I. It characterizes infectious
diseases like the common cold or influenza that do not provide immunity upon
infection. While the SIS model is too simplistic to explain the outbreak dynamics
of complex infectious diseases, it is the only compartment model with an explicit
analytical solution for the time course of its populations. This makes it a popular
model to explain and illustrate the basic principles of compartment modeling. The
learning objectives of this chapter on classical SIS modeling are to
• explain concepts of mass action incidence and constant rate recovery
• interpret contact and infectious rates and periods
• solve the analytical solutions for the susceptible and infectious populations
• distinguish disease free and endemic equilibria
• analyze the final size relation
• demonstrate how the infectious rate, reproduction number, and initial conditions
modulate outbreak dynamics
• discuss limitations of classical SIS modeling
By the end of the chapter, you will be able to analyze, simulate, and predict the
outbreak dynamics of simple infectious diseases like the common cold or influenza
that do not provide immunity to reinfection.
2.1 Introduction of the SIS model
Some infectious diseases, for example from the common cold or influenza, do not
provide lifelong immunity upon recovery from infection [2, 2], and previously in-
fected individuals become susceptible again [10]. We can model their behavior
through the simplest of all compartment models, the SIS model [5]. Figure 11.4
shows that the SIS model consists of only two populations, the susceptible group S
and the infectious group I [3]. As such, it provides valuable insight into the concept
33
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
E. Kuhl, Computational Epidemiology, https://doi.org/10.1007/978-3-030-82890-5_2](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-49-320.jpg)
![34 2 The classical SIS model
Fig. 2.1 Classical SIS model. The classical SIS model contains two compartments for the sus-
ceptible and infectious populations, S and I. The transition rates between the compartments, the
contact and infectious rates, β and γ, are inverses of the contact and infectious periods, B = 1/β
and C = 1/γ.
of compartment modeling [8]. The SIS model is often used to illustrate the basic
features of compartment models because we can solve the dynamics of its two pop-
ulations, S(t) and I(t), analytically at any point t in time [7]. The SIS model also has
simple analytical solutions for the converged final sizes, S∞ and I∞, as t → ∞ [13].
For the transition from the susceptible to the infectious group, we assume a mass
action incidence [4], which implies that the rate of new infections is proportional
to the size of the susceptible and infectious groups S and I weighted by the contact
rate β, Û
I = β SI. For the transition from the infectious to the susceptible group, we
Fig. 2.2 The transition from the susceptible to the infectious
group is based on the assumption of mass action incidence
for which the rate of new infections is proportional to the size
of the susceptible and infectious groups S and I, weighted
by the contact rate β, Û
I = β SI. The transition from the
infectious to the susceptible group is based on the assumption
of constant rate recovery for which the rate of recovered
infections is proportional to the size of the infectious group
I, weighted by the infectious rate γ, Û
I = −γI.
assume a constant rate recovery, which implies that the rate of recovered infections
is proportional to the size of the infectious group I weighted by the infectious rate
γ, Û
I = γI. Figure 2.2 illustrates the dynamics of these two assumptions that result in
the following system of two coupled ordinary differential equations,
Û
S = − β SI + γ I
Û
I = + β SI − γ I .
(2.1)
The transition rates between both compartments are the contact rate β and the
infectious rate γ in units [1/days], which are the inverses of the contact period
B = 1/β and the infectious period C = 1/γ in units [days]. The ratio between the
contact and infectious rates, or similarly, between the infectious and contact periods,
defines the basic reproduction number R0,
R0 =
β
γ
=
C
B
. (2.2)
In this simple format, the SIS model (2.1) neglects all vital dynamics, Û
S + Û
I 0 and
S + I = const. = 1. It does not account for births, natural deaths, or death from the
disease.](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-50-320.jpg)
![2.2 Analytical solution of the SIS model 35
2.2 Analytical solution of the SIS model
From the condition of non-vital dynamics, S = 1 − I, we can rephrase the system of
equations of the SIS model (2.1) in terms of only one independent variable, the size
of the infectious group I, governed by the following nonlinear ordinary differential
equation,
Û
I = β [ 1 − I ] I − γ I = [ β − γ ] I
1 −
I
1 − γ/β
. (2.3)
Equation (2.3) is a logistic differential equation of the form,
Û
I = r I [ 1 − I/K ] with r = β − γ and K = 1 − γ/β, (2.4)
where K is the carrying capacity. This type of equation has an explicit analytical
solution,
I(t) =
K I0
I0 + [ K − I0 ] exp (−r t)
, (2.5)
where I0 = I(0) is the initial infectious population [10]. It proves convenient to
reparameterize the equation for the infectious population (2.5) in terms of the basic
reproduction number R0 = β/γ and the infectious period C = 1/γ with r = [R0 −
1]/C and K = 1 − 1/R0. This provides the analytical solution for the SIS model in
terms of the infectious period C, the basic reproduction number R0, and the initial
infectious population I0,
S(t) = 1−
[1 − 1/R0] I0
I0 + [1 − 1/R0 − I0] exp ([1 − R0] t/C)
I(t) =
[1 − 1/R0] I0
I0 + [1 − 1/R0 − I0] exp ([1 − R0] t/C)
.
(2.6)
Figures 2.3, 2.4, and 2.5 highlight the outbreak dynamics of the SIS model, solved
analytically using equations (2.6), for the time period of one year. The three figures
demonstrate the sensitivity of the SIS model for varying infectious periods C, and
varying basic reproduction numbers R0, and initial infectious populations I0. Increas-
ing the infectious period C delays convergence to the endemic equilibrium, but the
final sizes S∞ and I∞ remain unchanged. Increasing the basic reproduction number
R0 accelerates convergence to the endemic equilibrium, decreases S∞, and increases
I∞. Increasing the initial infectious population I0 accelerates the onset of the out-
break, but the final sizes S∞ and I∞ remain unchanged. Interestingly, increasing the
initial exposed population I0 by an order of magnitude shifts the population dynam-
ics by a constant time increment, for the current parameterization by 50 days. This
highlights the exponential nature of the SIS model, which causes a constant acceler-
ation of the outbreak for a logarithmic increase of the initial infectious population,
while the overall outbreak dynamics remain the same.](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-51-320.jpg)
![36 2 The classical SIS model
Fig. 2.3 Classical SIS model. Sensitivity with respect to the infectious period C. Increasing
the infectious period C delays convergence to the endemic equilibrium, but the final sizes S∞ and
I∞ remain unchanged. Basic reproduction number R0 = 2.0, initial infectious population I0 = 0.01,
and infectious period C = 5, 10, 15, 20, 25, 30 days.
Fig. 2.4 Classical SIS model. Sensitivity with respect to the basic reproduction number R0.
Increasing the basic reproduction number R0 accelerates convergence to the endemic equilibrium,
decreases S∞, and increases I∞. Infectious period C = 20 days, initial infectious population
I0 = 0.01, and basic reproduction number R0 = 1.5, 1.7, 2.0, 2.4, 3.0, 5.0, 10.0.
2.3 Final size relation of the SIS model
For practical purposes, it is interesting to estimate the final susceptible and infectious
populations S∞ and I∞ [15]. From the analytical solution (2.5),
I∞ =
K I0
I0 + [ K − I0 ] exp (−r t∞)
with r = β − γ and K = 1 − γ/β, (2.7)](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-52-320.jpg)
![2.3 Final size relation of the SIS model 37
Fig. 2.5 Classical SIS model. Sensitivity with respect to the initial infectious population I0.
Increasing the initial infectious population I0 accelerates the onset of the outbreak, but the final
sizes S∞ and I∞ remain unchanged. Infectious period C = 20 days, basic reproduction number
R0 = 2.0, and initial infectious population I0 = 10−1, 10−2, 10−3, 10−4, 10−5, 10−6.
it is easy to show that we can distinguish two cases. If there is a finite initial infectious
population, I0 0, the infectious population will converge to either zero, for β γ,
r 0, and exp (−r t∞) → ∞, or to K for β γ, r 0, and exp (−r t∞) → 0 [2, 9],
thus
I∞ =
0 if β γ
K if β γ .
(2.8)
We can rephrase these limit conditions (2.8) in terms of the basic reproduction
number R0. In classical epidemiology, the two converged states are known as the
disease-free equilibrium and the endemic equilibrium [3]. The equations that define
the final converged populations of the endemic equilibrium state are the final size
relation.
R0 1 disease-free equilibrium: S∞ = 1 and I∞ = 0
R0 1 endemic equilibrium: S∞ = 1/R0 and I∞ = 1 − 1/R0 .
(2.9)
Figure 2.6 illustrates the final size relation and emphasizes the role of the basic
reproduction number R0 as the distinguishing outbreak characteristic between the
disease-free and endemic equilibrium. For reproduction numbers smaller than one,
R0 1, the SIS model converges to a disease-free equilibrium with S∞ = 1 and
I∞ = 0. For reproduction numbers larger than one, R0 1, the SIS model converges
to an endemic equilibrium with S∞ = 1/R0 and I∞ = 1 − 1/R0. The larger the
reproduction number R0, the smaller the final susceptible population S∞ and the
larger the final infectious population I∞ [3].](https://image.slidesharecdn.com/23120261-250602131124-c5d5c67b/85/Computational-Epidemiology-Datadriven-Modeling-Of-Covid19-Ellen-Kuhl-53-320.jpg)
Computational Epidemiology Datadriven Modeling Of Covid19 Ellen Kuhl
- 1. Computational Epidemiology Datadriven Modeling Of Covid19 Ellen Kuhl download https://ebookbell.com/product/computational-epidemiology- datadriven-modeling-of-covid19-ellen-kuhl-46240522 Explore and download more ebooks at ebookbell.com
- 2. Here are some recommended products that we believe you will be interested in. You can click the link to download. Computational Epidemiology Datadriven Modeling Of Covid19 Ellen Kuhl https://ebookbell.com/product/computational-epidemiology-datadriven- modeling-of-covid19-ellen-kuhl-46233584 Computational Epidemiology From Disease Transmission Modeling To Vaccination Decision Making 1st Ed Jiming Liu https://ebookbell.com/product/computational-epidemiology-from-disease- transmission-modeling-to-vaccination-decision-making-1st-ed-jiming- liu-22503076 Computational Intelligence And Data Analytics Proceedings Of Iccida 2022 1st Ed 2023 Rajkumar Buyya https://ebookbell.com/product/computational-intelligence-and-data- analytics-proceedings-of-iccida-2022-1st-ed-2023-rajkumar- buyya-45333718 Computational Methods In Organometallic Catalysis From Elementary Reactions To Mechanisms Yu Lan https://ebookbell.com/product/computational-methods-in-organometallic- catalysis-from-elementary-reactions-to-mechanisms-yu-lan-46084622
- 3. Computational Intelligence Based Solutions For Vision Systems Ansari Bajaj https://ebookbell.com/product/computational-intelligence-based- solutions-for-vision-systems-ansari-bajaj-46094270 Computational Methods Using Matlab An Introduction For Physicists P K Thiruvikraman https://ebookbell.com/product/computational-methods-using-matlab-an- introduction-for-physicists-p-k-thiruvikraman-46098316 Computational Semiotics Jeanguy Meunier https://ebookbell.com/product/computational-semiotics-jeanguy- meunier-46098550 Computational Engineering Of Historical Memories With A Showcase On Afroeurasia Ca 11001500 Ce Andrea Nanetti https://ebookbell.com/product/computational-engineering-of-historical- memories-with-a-showcase-on-afroeurasia-ca-11001500-ce-andrea- nanetti-46191608 Computational Mechanics With Neural Networks Genki Yagawa Atsuya Oishi https://ebookbell.com/product/computational-mechanics-with-neural- networks-genki-yagawa-atsuya-oishi-46245306
- 5. Computational Epidemiology Ellen Kuhl Data-Driven Modeling of COVID-19
- 7. Ellen Kuhl Computational Epidemiology Data-Driven Modeling of COVID-19
- 8. Ellen Kuhl Mechanical Engineering Stanford University Stanford, CA, USA ISBN 978-3-030-82889-9 ISBN 978-3-030-82890-5 (eBook) https://doi.org/10.1007/978-3-030-82890-5 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors, and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
- 9. To my mom, with whom I had the most scientific discussions about COVID-19.
- 10. Foreword Nothing will ever be quite the same. The Great COVID-19 Pandemic that started in December 2019 will be remem- bered as the main event of the first part of the 21st Century. It affected all parts of the world, it laid bare social inequalities, revealed the excesses of modern life, shook all strata of society, and changed profoundly how individuals, groups, and nations interact with each other. Early in the crisis, it was realized that theoretical epidemiology and modeling are essential tools to confront the problem. As Daniel Bernoulli stated in the first-ever epidemiological studies of 1760: "I simply wish that, in a matter which so closely concerns the wellbeing of the human race, no decision shall be made without all the knowledge which a little analysis and calculation can provide". By January 2020, the first ‘little analysis and calculation’ provided by epidemiological studies revealed the seriousness of the situation and scientists around the world rang the alarm bell. As the crisis escalated, the attention of the media, governments, and the public turned towards modelers with burning questions: how bad is it? how will it evolve? how many cases should we anticipate? will the hospitals be overwhelmed? how many deaths? what can we do? Like many people in academia, Ellen Kuhl found herself naturally interested in the problem. As an extreme athlete, a marathoner, a triathlete and an iron-(wo)man, Ellen is fearless. Early on, she realized that the methods she had been developing to understand neurodegenerative diseases could be readily adapted to the evolving crisis. As a scientist, Ellen combines a unique ability for modeling, great technical skills, and a wonderful intuition for good problems. She quickly built elegant data- driven models for how the disease spreads around different parts of the world through the airline network that soon became landmark studies. As an educator, she also realized that students in Stanford would gain from learning more about the science behind the disease. Rapidly, she put together a new course that was taught in the 2020 winter term, as the second wave was gathering strength. In her course that I attended and enjoyed tremendously, she taught both the basics of epidemiology and her own research as well as many topics related to the wider social and political vii
- 11. viii Foreword impacts of the crisis. This book is the combined result of her course, state-of-the-art research, and general reflections. The first part of the book is a fast-paced self-contained review of the fundamental ideas of epidemiology. It is full of wonderful stories, biographies, and anecdotes that brings life to the topic. It introduces and defines important concepts such as the reproduction number, herd immunity, and vaccine efficacy. On the mathematical side, it is centered around the development of the the famous SIR model that captures the evolution of three homogeneous populations of people susceptible (S ), Infected (Is ) and Recovered (R) going back to the work of Kermack and McKendrick in 1927. The model will tell you that for a reproductive number R larger than one, after an initial exponential growth, there will be saturation when enough of the population has been infected before the disease eventually runs its course. This classic material is updated for our computational world in Part II where Ellen shows how to implement efficiently the classic theory and how it can be generalized to problems that are directly relevant to the COVID crisis. In particular, in Chapter 7 and 8 two extra populations, the exposed (E) and asymptomatic (Ia), are introduced to capture the propagation of the disease through populations that do not display yet signs of infections, a crucial feature of this disease. As Ellen shows, this simple system of four or five differential equations is already enough to capture many features of the disease and can be used to understand early disease outbreaks. The third part of the book brings the topic of epidemiology to current research level. One of the key features of a pandemic is the propagation of the disease from one region to another. In our hyper-connected world there are many ways diseases can travel. At the global level, the airline network is a natural way for the disease to spread especially through people who show no symptoms. Ellen shows how the SEIR model can then be extended to a network and efficiently implemented. Remarkably, this simple idea is sufficient to understand the early spread and different phases of the disease. Yet, a key question remains. How do we use daily data about the number of cases to validate and infer parameters for the various models? Part IV of the book addresses directly this question and brings the topic to the forefront of current data research by presenting Bayesian inference methods and how it can be efficiently applied to the COVID crisis. This data-driven approach combined with the network models bring state-of-the-art techniques to the study of the disease dynamics and shows how secondary data, such as mobility data from cell phones, can be used as a barometer to predict the development of the disease. Through the fog of war, governments have made many mistakes that have cost countless lives. As Anne Frank wrote: “What is done cannot be undone, but at least one can keep it from happening again". How do we keep it from happening? We learn from our mistakes, sharpen our tools and models and confront the current and next crises with the full power of science. Ellen’s wonderful book gives us the ideas, methods, tools, knowledge, and concepts to understand the current pandemic and be prepared for the next one. Oxford, May 2021 Alain Goriely
- 12. Preface The objective of this book is to understand the outbreak dynamics of the COVID-19 pandemic through the lens of computational modeling. Computational modeling can provide valuable insights into the dynamics and control of a global pandemic to guide public health decisions. So... why didn’t it? Why did dozens of computational models make predictions that were orders of magnitude off? This book seeks to an- swer this question by integrating innovative concepts of mathematical epidemiology, computational modeling, physics-based simulation, and probabilistic programming. We illustrate how we can infer critical disease parameters–in real time–from reported case data to make informed predictions and guide political decision making. We crit- ically discuss questions that COVID-19 models can and cannot answer and showcase controversial decisions around the early outbreak dynamics, outbreak control, and gradual return to normal. As scientists, it is our ethic responsibility to educate the public to ask the right questions and to communicate the limitations of our answers. Throughout this book, we will create data-driven models for COVID-19 to do so. Who is this book for? If you are a student, educator, basic scientist, or medical researcher in the natural or social sciences, or someone passionate about big data and human health: This book is for you! Don’t worry, this book is introductory and doesn’t require a deep knowledge in epidemiology. A fascination for numbers and a general excitement for physics-based modeling, data science, and public health are a lot more important. And this is why this book is both, a textbook for undergraduates and graduate students, and a monograph for researchers and scientists. As a textbook, this book can be used in the mathematical life sciences suitable for courses in applied mathematics, biomedical engineering, biostatistics, computer science, data science, epidemiology, health sciences, machine learning, mathematical biology, numerical methods, and probabilistic programming. As a monograph, this book integrates the basic fundamentals of mathematical and computational epidemiology with modern concepts of data-driven modeling and probabilistic programming. It serves researchers in epidemiology and public health, with timely examples of computational modeling, and scientists in the data science and machine learning, with applications to COVID-19 and human health. ix
- 13. x Preface What does this book cover? This book consists of four main parts that gradually build from mathematical epidemiology via computational and network epidemiology to data-driven epidemiology. Part I. Mathematical Epidemiology introduces the basic concepts of epidemiology in view of the COVID-19 pandemic with the objectives to understand the cause of an infectious disease, predict its outbreak dynamics, and design strategies to control it. We introduce the paradigm of compartment modeling and revisit analytical solutions for the classical SIS, SIR, and SEIR models using outbreak data of the COVID-19 pandemic. Typical problems include estimating the herd immunity threshold and efficacy of different COVID-19 vaccines and the growth rate, basic reproduction number, contact period of the COVID-19 outbreak at different locations. Part II. Computational Epidemiology introduces numerical methods for ordinary differential equations and applies these methods to discretize, linearize, and solve the governing equations of SIS, SIR, SEIR, and SEIIR models to interpret the case data of COVID-19. Typical problems include simulating the first wave of the COVID-19 outbreak using reported case data and understanding the effects of early community spreading and outbreak control. We discuss why classical epidemiology models have failed to predict the outbreak dynamics of COVID-19 and introduce new concepts of data-driven dynamic contact rates and serology-informed asymptomatic transmission to address these shortcomings. Part III. Network Epidemiology discusses numerical methods for partial differ- ential equations and applies these methods to discretize, linearize, and solve the spreading of infectious diseases using discrete mobility networks and finite element models. Instead of solving the outbreak dynamics of COVID-19 locally for each region, state, or country, we now allow individuals to travel and populations to mix globally, informed by cell phone mobility data and air travel statistics. Typi- cal problems include understanding the early outbreak patterns, the effect of travel restrictions, and the risk of reopening after lockdown. Part IV. Data-driven Epidemiology covers the most timely methods and applica- tions of this book. It focuses on probabilistic programming with the objectives to understand, predict, and control the outbreak dynamics of the COVID-19 pandemic. We integrate computational epidemiology and data-driven modeling to explore dis- ease data in view of different compartment models using a probabilistic approach and quantify the uncertainties of our analysis. Typical problems include inferring the reproduction dynamics, visualizing the effects of asymptomatic transmission, and correlating case data and mobility. This book is by no means complete. It does not cover agent-based modeling, age- dependent modeling, population mixing, purely statistical, stochastic, or probabilistic modeling, forecasting, and many other key aspects of computational epidemiology. Instead, it is a personal reflection on the role of data-driven modeling during the COVID-19 pandemic, motivated by the curiosity to understand it. Because Science. Stanford, May 2021 Ellen Kuhl
- 14. Acknowledgements Now, that’s a wrap! I started writing this book on January 14, 2021, during the United States peak of the COVID-19 pandemic and finished on May 31, 2021, at the lowest incidence in 14 months. Dozens of people have contributed to this book, directly or indirectly, through endless discussions around the pandemic. I was fortunate to collaborate with an amazing international team, Alain Goriely from the UK, Francisco Sahli Costabal from Chile, Henry van den Bedem from the Netherlands, Kevin Linka from Germany, Mathias Peirlinck from Belgium, Paris Perdikaris from Greece, and Proton Rahman from Canada. The daily discussions of our local outbreak dynamics–while never meeting in person–were an essential part of my pandemic life made possible by Zoom. Thank you for sharing this unique experience! I enjoyed regular COVID-19 meetings with my scientific friends around the world, Silvia Budday, Krishna Garikipati, Tom Hughes, Tinsley Oden, Paul Steinmann, Tarek Zohdi, and the IMAG/MSM Working Group for Multiscale Modeling and Viral Pandemics. But the true motivation for this book came from the students of the new course Data-driven modeling of COVID-19 that we created at Stanford University in the Fall of 2020 to make online learning a bit more tangible. The enrollment ranged from undergraduates in biology, classics, computer science, engineering, ethics, human biology, mechanical engineering, and Spanish, to master and PhD students in aeronautics, astronautics, computer science, environment and resources, management science and engineering, and mechanical engineering. Together, we designed this course as the pandemic unfolded–in real time–and this book is a collection of class notes and feedback from students, guest speakers, and lecturers. We received support from the Stanford Bio-X Program and the School of Engineering COVID-19 Research and Assistance Fund. Massive thanks to all students in the class, to our amazing course assistants, Amelie Schäfer, Oguz Tikenogullari, Mathias Peirlinck, and Kevin Linka, and to my favorite guest speaker, Alain Goriely. Last but not least, I thank the true heroes of the pandemic, Jasper and Syb, and all the kids who not only had to bear the uncertainties of a global pandemic, but also endure their parents working from home. Thank you, Henry, for getting through this together, with 354 miles of swimming, 3611 miles of biking, and 3279 miles of running. This has been a truly memorable time. I’m glad it’s over–at least for now. xi
- 15. Contents Part I Mathematical epidemiology 1 Introduction to mathematical epidemiology . . . . . . . . . . . . . . . . . . . . . . . 3 1.1 A brief history of infectious diseases . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.2 Introduction to epidemiology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 1.3 Testing, testing, testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 1.4 The basic reproduction number . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 1.5 Concept of herd immunity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 1.6 Concept of immunization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 1.7 Mathematical modeling in epidemiology . . . . . . . . . . . . . . . . . . . . . . . 19 1.8 Data-driven modeling in epidemiology. . . . . . . . . . . . . . . . . . . . . . . . . 23 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 2 The classical SIS model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 2.1 Introduction of the SIS model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 2.2 Analytical solution of the SIS model. . . . . . . . . . . . . . . . . . . . . . . . . . . 35 2.3 Final size relation of the SIS model. . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 3 The classical SIR model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 3.1 Introduction of the SIR model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 3.2 Growth rate of the SIR model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 3.3 Maximum infection of the SIR model . . . . . . . . . . . . . . . . . . . . . . . . . . 46 3.4 Final size relation of the SIR model . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 3.5 The Kermack-McKendrick theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 xiii
- 16. xiv Contents 4 The classical SEIR model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 4.1 Introduction of the SEIR model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 4.2 Growth rate of the SEIR model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 4.3 Maximum exposure and infection of the SEIR model . . . . . . . . . . . . . 66 4.4 Final size relation of the SEIR model . . . . . . . . . . . . . . . . . . . . . . . . . . 68 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 Part II Computational epidemiology 5 Introduction to computational epidemiology . . . . . . . . . . . . . . . . . . . . . . 81 5.1 Numerical methods for ordinary differential equations . . . . . . . . . . . . 81 5.2 Explicit time integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 5.3 Implicit time integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 5.4 Comparison of explicit and implicit time integration . . . . . . . . . . . . . 92 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 6 The computational SIR model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 6.1 Explicit time integration of the SIR model . . . . . . . . . . . . . . . . . . . . . . 99 6.2 Implicit time integration of the SIR model . . . . . . . . . . . . . . . . . . . . . . 101 6.3 Comparison of explicit and implicit SIR models . . . . . . . . . . . . . . . . . 102 6.4 Comparison of SIR and SIS models . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 6.5 Sensitivity analysis of the SIR model . . . . . . . . . . . . . . . . . . . . . . . . . . 105 6.6 Example: Early COVID-19 outbreak in Austria . . . . . . . . . . . . . . . . . . 106 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 7 The computational SEIR model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 7.1 Explicit time integration of the SEIR model . . . . . . . . . . . . . . . . . . . . 117 7.2 Implicit time integration of the SEIR model . . . . . . . . . . . . . . . . . . . . 119 7.3 Comparison of explicit and implicit SEIR models. . . . . . . . . . . . . . . . 121 7.4 Comparison of SEIR and SIR models . . . . . . . . . . . . . . . . . . . . . . . . . . 124 7.5 Sensitivity analysis of the SEIR model . . . . . . . . . . . . . . . . . . . . . . . . . 124 7.6 Maximum exposure and infection of the SEIR model . . . . . . . . . . . . . 127 7.7 Dynamic SEIR model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 7.8 Example: Early COVID-19 outbreak dynamics in China . . . . . . . . . . 131 7.9 Example: Early COVID-19 outbreak control in Germany . . . . . . . . . 136 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 8 The computational SEIIR model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 8.1 Introduction of the SEIIR model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 8.2 Discrete SEIIR model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 8.3 Similar symptomatic and asymptomatic disease dynamics. . . . . . . . . 153 8.4 Different symptomatic and asymptomatic disease dynamics . . . . . . . 155
- 17. Contents xv 8.5 Example: Early COVID-19 outbreak in the Netherlands. . . . . . . . . . . 158 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164 Part III Network epidemiology 9 Introduction to network epidemiology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169 9.1 Numerical methods for partial differential equations. . . . . . . . . . . . . . 169 9.2 Network diffusion method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 9.3 Finite element method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180 9.4 Comparison of network diffusion and finite element methods . . . . . . 191 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198 10 The network SEIR model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199 10.1 Network diffusion method of the SEIR model . . . . . . . . . . . . . . . . . . . 199 10.2 Example: COVID-19 spreading across the United States . . . . . . . . . . 202 10.3 Example: Travel restrictions across the European Union . . . . . . . . . . 207 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216 Part IV Data-driven epidemiology 11 Introduction to data-driven epidemiology . . . . . . . . . . . . . . . . . . . . . . . . . 221 11.1 Introduction to data-driven modeling . . . . . . . . . . . . . . . . . . . . . . . . . . 222 11.2 Bayes’ theorem for discrete parameter values . . . . . . . . . . . . . . . . . . . 225 11.3 Bayes’ theorem for continuous parameter values . . . . . . . . . . . . . . . . . 228 11.4 Data-driven SIS model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247 12 Data-driven dynamic SEIR model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249 12.1 Introduction of the data-driven dynamic SEIR model . . . . . . . . . . . . . 250 12.2 Example: Inferring reproduction of COVID-19 in Europe . . . . . . . . . . 256 12.3 Example: Correlating mobility and reproduction of COVID-19 . . . . . 259 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266 13 Data-driven dynamic SEIIR model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269 13.1 Introduction of the data-driven dynamic SEIIR model . . . . . . . . . . . . 270 13.2 Example: Visualizing asymptomatic COVID-19 transmission . . . . . . 277 13.3 Example: Asymptomatic COVID-19 transmission worldwide. . . . . . . 280 13.4 Example: Inferring the outbreak date of COVID-19 . . . . . . . . . . . . . . 282 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
- 18. xvi Contents 14 Data-driven network SEIR model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289 14.1 Introduction of the data-driven network SEIR model . . . . . . . . . . . . . 289 14.2 Example: The Newfoundland story . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
- 20. Chapter 1 Introduction to mathematical epidemiology Abstract Mathematical epidemiology is the science of understanding the cause of a disease, predicting its outbreak dynamics, and developing strategies to control it. Throughout the past century, mathematical epidemiology has become the corner- stone of public health. This success story is a result of the rapid advancement of mathematical models and the broad availability of massive amounts of data. Now, two fundamental things have changed that may forever affect our understanding of mod- ern epidemiology: the outbreak of the COVID-19 pandemic, in the age of machine learning. Here, to embrace this opportunity, we introduce the concept of mathe- matical epidemiology in light of data-driven modeling and physics-based learning. We briefly review the traditional concepts of epidemiology, and discuss how our classical understanding of testing, reproduction, herd immunity, and immunization has changed in view of the COVID-19 pandemic. We compare the strengths and weaknesses of purely statistical and mechanistic models and illustrate how we can integrate the large volume of COVID-19 data into mechanistic compartment models to infer model parameters, learn correlations, and identify causation. The learning objectives of this chapter on mathematical epidemiology are to • recognize the ever-present risk of infectious diseases, now and beyond COVID-19 • interpret epidemiological data from molecular, antigen, and antibody tests • explain the basic reproduction number and its impact on public health decisions • estimate herd immunity thresholds for given basic reproduction numbers • demonstrate how vaccination can accelerate the path towards herd immunity • calculate the efficacy and risk ratio of a vaccine • understand the strengths and weaknesses of statistical and mechanistic models • explain the importance of nonlinear feedback in compartment modeling • find and prepare COVID-19 case data towards creating data-driven models • discuss limitations of testing, reported data, and modeling By the end of the chapter, you will be able to think in terms of data-driven modeling, know where to find and how to interpret your data, judge how to select your model, and understand why so many models have failed to explain and predict the outbreak dynamics of the COVID-19 pandemic. 3 © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 E. Kuhl, Computational Epidemiology, https://doi.org/10.1007/978-3-030-82890-5_1
- 21. 4 1 Introduction to mathematical epidemiology 1.1 A brief history of infectious diseases In the year 2020, everybody–at home on their couch–became an infectious disease expert. To calibrate all this newly gained knowledge, it seems a good idea to briefly reiterate the basic concepts and nomenclature of infectious diseases. Infectious dis- eases are spread by either bacterial or viral agents and are ever-present in society [8]. Every once in a while, this may result in outbreaks that have a significant impact on a local or global level. Depending on their spatial and temporal spread, we can classify outbreaks as endemic, epidemic, or pandemic [28]. Endemic outbreaks, for example chickenpox, are permanently present in a region or a population. Epidemic outbreaks, for example the seasonal flu, affects a lot of people in a short period of time, spread across several communities, and then disappear. Pandemic outbreaks, for example the Spanish flu, are epidemic outbreaks that affects a lot of people in a short period of time and spread across the entire world. Mathematical modeling of infectious diseases is important to understand their outbreak dynamics and inform political decision making to manage their spread [4]. Table 1.1 History of recent infectious disease outbreaks. Time period, type of disease, number of deaths, and location. period disease deaths location 1346 - 1350 Black Death 100,000,000 1/3 of Europe 1665 - 1666 Great Plague 100,000 1/4 of London 1918 - 1920 Spanish flu 50,000,000 worldwide 1980 measles 2,600,000 worldwide / year 2003 SARS 774 worldwide 2009 - 2010 H1N1 18,500 worldwide 2011 tuberculosis 1,400,000 worldwide / year HIV/AIDS 1,200,000 worldwide / year malaria 627,000 worldwide / year 2011 measles 160,000 worldwide (-94%) 2012 - 2020 MERS 866 worldwide seasonal flu 35,000 United States / year 2014 - 2016 ebola 11,000 Africa 2018 ebola 2,280 Congo 2019 -2021 COVID-19 3,300,000 worldwide The COVID-19 pandemic. On March 11, 2020, Tedros Adhanom Ghebreyesus, the Director-General of the World Health Organization, declared the COVID-19 outbreak a global pandemic [49]. On that day, it had affected 126,702 people worldwide. What nobody could have foreseen is that within the following year, by March 11, 2021, this number had increased by three orders of magnitude, to 118.57 million [12]. Naturally, we often think of the COVID-19 pandemic as the deadliest and most devastating infectious disease in modern history.
- 22. 1.1 A brief history of infectious diseases 5 Endemic. The word endemic is derived from the greek words en meaning in and demos meaning people. An infectious disease is endemic when it is constantly maintained at a baseline level in a geographic region without external inputs. For example, chicken pox is endemic in the United Kingdom. A person-to-person transmitted disease is endemic if each infected person passes the disease to one other person on average. The infection neither dies out nor increases exponentially, it is in an endemic steady state. Infectious disease experts ask: • How many people are infectious at any give time? – What is I(t)? • How fast do new infections arise? – What is dI/dt? – Stability analysis • What are the effects of quarantine or vaccination? – What is R0? • Can we eradicate the disease? – What is H = 1 − 1/R0? – Limit analysis Epidemic. The word epidemic is derived from the greek words epi meaning upon and demos meaning people. An infectious disease is epidemic when it spreads rapidly across a large number of people in a short time period. For example, the seasonal flu is epidemic. Epidemics often come in waves with several recurring outbreaks, which can sometimes be seasonal. There is usually no increase in susceptibles and the disease dies out as the number of infectives decreases because a large enough fraction of the population has become immune. Infectious disease experts ask: • How severe will the epidemic be? – What are Imax? • When will it reach its peak? – What is t(Imax)? • How long will it last? – What are S∞ and R∞? What is t(S∞)? • What are the effects of vaccination? – What are R0 and H = 1−1/R0 and I? Pandemic. The word pandemic is derived from the greek words pan meaning all and demos meaning people. An infectious disease is pandemic when it spreads rapidly across a large region or worldwide in a short time period. A pandemic is a global outbreak. For example, smallpox, tuberculosis, the black death, the Spanish flu were pandemics, and COVID-19 is now. Infectious disease experts ask: • How severe is the pandemic, when will it peak? – What are Imax and t(Imax)? • What are effective measures to manage the outbreak? – What is R(t)? • What are the effects of quarantine or lockdown? – What is β(t)? • What are the effects of travel restrictions? – What are κ and Lij? • How do we prioritize vaccination? – What are R0 and H = 1 − 1/R0 and I?
- 23. 6 1 Introduction to mathematical epidemiology Table 1.1 summarizes recent infectious disease outbreaks ranging from the Black Death in the 14th century with an estimated 100 million deaths across Europe to the current COVID-19 pandemic with 3.3 million deaths worldwide to date. Notably, in 1980, the annual death toll of the measles with 2.6 million was of comparable size. Today, after massive vaccination campaigns, this number has dropped significantly, by 94% to 160,000 [5]. Fig. 1.1 The Lessons of the Pandemic. In this famous Science publication, George A. Soper reflects on the Spanish flu that resulted in 50 million deaths worldwide in 1918 and 1919. Figure 1.1 shows the cover page of the Science publication The Lessons from the Pandemic from May 1919, a reflection on the scientific understanding of the Spanish flu [29]. While some of it is specific to the nature of the 1918 influenza and the time during which it occurred, much of it still applies to the COVID-19 pandemic today: The Spanish flu lasted from 1918 to 1920 and, similar to COVID-19, occurred in multiple waves. Both, the Spanish flu and COVID-19, are contagious respiratory illnesses that spread from person to person, mainly by droplets, through cough, sneeze, or talk. Naturally, increased hygiene, mask wearing, and physical distancing in addition to strict isolation and quarantine are successful strategies to manage both conditions [10]. While COVID-19 is caused by a coronavirus, SARS-CoV-2, the 1918 influenza was caused by the H1N1 influenza A virus, a virus of avian origin. Within 25 months, it infected 500 million people, one third of the world’s population, and resulted in more than 50 million deaths [5]. The major differences between the Spanish flu and COVID-19 are their high risk populations and the mechanisms of death: In contrast to the Spanish flu, which affected mainly healthy adults between 25 and 40 years of age, COVID-19 affects mainly individuals of 65 years and older with comorbidities. Victims of the 1918 influenza mainly died from secondary bacterial pneumonia, whereas victims of COVID-19 die from an overactive immune response that results in organ failure [32]. Nevertheless, comparing the COVID- 19 outbreak to previous pandemics can provide insight and guidance to manage the current COVID-19 pandemic. An important element of this comparison are mathematical and computational tools that have been designed to quantify and explain the spreading mechanisms of infectious diseases. This is the objective of mathematical and computational epidemiology [3].
- 24. 1.2 Introduction to epidemiology 7 1.2 Introduction to epidemiology Epidemiology is the study of distributions, patterns, and determinants of health- related events in human populations [3]. It is a cornerstone of public health , and shapes policy decisions by identifying risk factors for outbreaks and targets for prevention [28]. Epidemiology literally means the study of what is upon the people. It is derived from the greek words epi, meaning upon demos, meaning people, and logos meaning study, suggesting that it applies only to human populations. By this definition, epidemiology is the scientific, systematic, data-driven study of the distribution, i.e., where, who, and when, and determinants, i.e., causes and risk factors, of health-related patterns and events in specified populations, i.e., the world, a country, state, county, city, school, neighborhood. Epidemiology also includes the study of outbreak dynamics, outbreak control, and informing political decision making. Agent, host, and environment. A key premise in epidemiology is that health-related events are not evenly distributed in a population; rather, they affect some individuals more than others [3]. An important goal of epidemiology is to identify the causes that put these individuals at a higher risk [2]. A simple but popular model to analyze and explain disease causation is the epidemiologic triangle. The epidemiologic triangle summarizes the interplay of the three components that contribute to the spread of an infectious disease: an external agent, a susceptible host, and an environment in which agent and host interact [9]. Descriptive epidemiologists characterize this interaction as the seed, the soil, and the climate [42]. Effective public health measures assess all three components and their interactions to control or prevent the spreading of a disease. Figure 1.2 illustrates the epidemiologic triangle of COVID-19 with agent, host, and environment. For COVID-19, the agent is the SARS-CoV-2 virus, the hosts are people, and the environment are droplets. Interventions between any two of these three components can help reduce the spread of the disease [20]. For exam- ple, reduced exposure, vaccination, and antiviral treatment can modulate agent-host interactions and reduce the number of new infections. From descriptive to mathematical epidemiology. Mathematical models of infec- tious diseases date back to Daniel Bernoulli’s model for smallpox in 1760 [5], and they have been developed and improved extensively since the 1920s [14]. In the middle of the 19th century, the English physician John Snow conducted a famous series of experiments of the cholera outbreak in London to discover the cause of the disease and to prevent its recurrence [46]. Because his research illustrates the classic sequence–from descriptive epidemiology and hypothesis generation, to mathemat- ical epidemiology and hypothesis testing–John Snow is considered the father of modern epidemiology. Figure 1.4 summarizes the experiments of John Snow that ended the cholera outbreak in London by removing the handle of a public water
- 25. 8 1 Introduction to mathematical epidemiology Fig. 1.2 Epidemiologic triangle of COVID-19 with agent, host, and environment. The epidemi- ologic triangle illustrates the interplay of the three components that contribute to the spread of a disease: an external agent, a susceptible host, and an environment in which agent and host interact. For COVID-19, the agent is the SARS-CoV-2 virus, the hosts are people, and the environment are droplets. Interventions between any of these three components can reduce the spread of the disease. Fig. 1.3 Daniel Bernoulli is considered the most famous mathematician of the Bernoulli family, although he actually studied medicine. He was born on February 8, 1700 in Groningen, the Netherlands, and is best known for his applications of mathematics to mechanics. To prove the efficacy of vaccination against smallpox, he proposed the first compartment model in 1760 [5]. It only had two compartments, but demonstrated the potential of modeling to understand the mechanisms of transmission, predict future spread, and control the outbreak through vaccination. Daniel Bernoulli died on March 27, 1782 in Basel, Switzerland. pump. Today the field of epidemiology is considered a quantitative discipline that uses rigorous mathematical tools of probability and statistics to develop and test hypotheses to understand and explain health-related events [3]. In simple terms, today’s epidemiologists count, scale, and compare. They count the number of cases; scale this count by a characteristic population to define fractions; and compare the evolution of these fractions over time [9]. Mathematical epidemiology has advanced significantly throughout the past decades [11], and models have become more so- phisticated and complex [39]. This implies that in epidemiology today, is often no longer possible to solve epidemiological models analytically and in closed form [3]. From mathematical to computational epidemiology. Computational epidemiol- ogy is a multidisciplinary field that integrates mathematics, computer science, and public health to better understand central questions in epidemiology such as the
- 26. 1.2 Introduction to epidemiology 9 Fig. 1.4 The English physician John Snow is considered one of the founders of modern epidemiology, in part because of his research during the cholera outbreak in London in 1954 [46]. John Snow lived from March 15, 1813 to June 16, 1858. He questioned the widely accepted paradigm that cholera would spread by polluted bad air. He postulated that water would be the cause of the cholera outbreak in Soho, London in 1854. By talking to local residents, he identified the source of the outbreak as the public water pump on Broad Street. Although his chemical and microscopic observations were inconclusive, his spreading patterns of the cholera outbreak were convincing enough to persuade the local council to disable a street pump on Broad Street by removing its handle. There is a common belief that this action marked the ending of the outbreak. spread of diseases or the effectiveness of public health interventions [22]. Real-time epidemiology is a rapidly developing area within computational epidemiology that seeks to support policy makers–in real time–as an outbreak is unfolding. An impor- tant application of real-time epidemiology is disease surveillance, the data-driven collection, analysis, and interpretation of large volumes of disease data from a va- riety of sources. This information can help to evaluate the effectiveness of control and preventative health measures [12]. Central to computational epidemiology is the accurate knowledge of people who are affected by an infectious disease at any given point in time. During the COVID-19 pandemic, for the first time in history, this information has been collected conscientiously, shared publicly, updated daily, and made freely available in real time [13]. Sensitivity and specificity. A diagnostic tests, for example, the nasal swab test, can be inaccurate in two ways: a false positive result erroneously labels a healthy person as infected, resulting in unnecessary quarantine and contact tracing; a false negative result oversees an infected person, resulting in the risk of infecting others. sensitivity = true positives / all sick people Sensitivity measures the proportion of positives, diseased people, that are correctly identified and not overlooked. specificity = true negatives / all healthy people Specificity measures the proportion of negatives, healthy people, that are correctly identified and not classified as diseased. A perfect test has 100% sensitivity and 100% specificity.
- 27. 10 1 Introduction to mathematical epidemiology 1.3 Testing, testing, testing During a global pandemic, knowing how many people currently have the disease or have previously had it provides crucial information for healthcare providers and policy makers, both on the individual and population levels. Unfortunately, there is no method to provide this information with absolute accuracy. We usually use two measures to characterize the degree of accuracy of a test: sensitivity, the fraction of correctly identified positive, diseased individuals; and specificity, the fraction of correctly identified negative, healthy individuals. Understanding these limitations is important when modeling, simulating, and predicting the outbreak dynamics of a pandemic, especially in view of disease management and political decision making. Table 1.2 Testing for COVID-19. Summary of the three most common tests for SARS-CoV-2. Diagnostic tests, including molecular and antigen tests, provide information about acute infection, whereas antibody or serology tests provide information about previous infection. Publicly available case data are based on diagnostic testing; seroprevalence reports are based on antibody testing. diagnostic testing antibody testing molecular testing antigen testing serology testing provides information about acute COVID-19 infection provides information about acute COVID-19 infection does not provide information about acute infection does not provide information about past infection does not provide information about past infection tests for antibody presence, does not guarantee immunity can take hours to days takes minutes to hours takes minutes to hours relatively accurate early in the infection faster and cheaper than molecular tests quick results accuracy drops later in infec- tion less accurate than molecular tests tests can vary in accuracy point of care testing; collect swabs at home and mail in point of care testing; collect swabs at home and mail in blood test that can be done at doctor’s office or at home Testing for COVID-19. Since its outbreak in late 2019, the COVID-19 pandemic has generated an exponentially growing demand for testing, and diagnostic assays that enable mass screening have been developed at an unprecedented pace. To success- fully use these tests to inform public health strategies, it is critical to understand their individual strengths and limitations. There are two different types of assays for SARS-CoV-2, the virus that causes COVID-19: diagnostic tests and antibody tests. A diagnostic test can show if you have an active COVID-19 infection and need to take steps to quarantine or isolate yourself from others. Molecular tests and antigen tests fall under this category. A molecular test is a diagnostic test that detects genetic material from the virus using, for example, reverse transcription polymerase chain reaction or nucleic acid amplification. An antigen test is a diagnostic test that detects specific proteins made by the virus. Samples for diagnostic tests are typically col- lected with a nasal or throat swab or with saliva from spitting into a tube. Diagnostic testing is critical to provide early treatment, quarantine individuals sooner, and trace and isolate their contacts to reduce the spread of the virus.
- 28. 1.4 The basic reproduction number 11 Fig. 1.5 Testing for COVID-19. The objective of COVID-19 testing is to probe whether an individual is currently or has previously been infected with SARS-CoV-2. A nasal or throat swab test is a diagnostic test that provides information about an acute infection. Diagnostic testing is critical to provide early treatment, quarantine individuals, and trace and isolate their contacts to reduce the spread of the virus. A blood test is an antibody test that provides information about a previous infection. Antibody tests are critical to estimate the overall dimension of an outbreak. Throughout this book, we use COVID-19 case data from public dashboards based on confirmed positive diagnostic tests. An antibody test can detect antibodies that are made by your immune system in response to a previous infection with SARS-CoV-2. In contrast to genetic material from the virus or specific proteins made by the virus, antibodies can take several days or even weeks to develop, but they can remain present for several weeks or months after recovery. A serology test is an antibody test that looks for antibodies in blood samples. These venous blood samples are typically collected at the doctor’s office or in the clinic. Antibody tests should not be used to diagnose an active infec- tion. Instead, antibody testing is important to understand the kinetics of the immune response to infection, clarify whether an infection protects from future infection, characterize how long immunity will last, and estimate the overall dimension of an outbreak. Table 1.2 contrasts the three most common types of tests for COVID-19. Throughout this book, we use COVID-19 case data from public dashboards based on confirmed positive diagnostic tests. Only in Chapters 8 and 13, where we explore the effects of asymptomatic transmission, we use specialized models that combine case data from diagnostic tests with seroprevalence data from antibody tests. 1.4 The basic reproduction number The basic reproduction number is a powerful but simple concept to explain the contagiousness and transmissibility of an infectious disease [6]. For decades, epi- demiologists have successfully used the basic reproduction number to quantify how many new infections a single infectious individual creates in an otherwise com- pletely susceptible population [6]. During the COVID-19 pandemic, the public me- dia, scientists, and political decision makers across the globe have adopted the basic reproduction number as an illustrative metric to explain and justify the need for dif- ferent outbreak control strategies [10]: An outbreak will continue for reproduction numbers larger than one, R0 > 1, and come to an end for reproduction numbers smaller than one, R0 < 1 [8]. However, especially in the midst of a global pandemic, it is difficult–if not impossible–to measure R0 directly [48].
- 29. 12 1 Introduction to mathematical epidemiology Table 1.3 Basic reproduction numbers and herd immunity thresholds for common infectious diseases. Herd immunity is the indirect protection from an infectious disease that occurs when a large fraction of the population has become immune. The herd immunity threshold, H = 1 − 1/R0, beyond which this protection occurs is a function of the basic reproduction number R0. disease R0 H disease R0 H measles 12 - 18 92 - 95% mumps 4.0 - 7.0 75 - 86% pertussis 12 - 17 92 - 94% COVID-19 2.0 - 6.0 50 - 83% rubella 6 - 7 83 - 86% SARS 2.0 - 5.0 50 - 80% smallpox 6 - 7 83 - 86% ebola 1.5 - 2.5 33 - 60% polio 5 - 7 80 - 86% influenza 1.5 - 1.8 33 - 44% Table 1.3 summarizes the basic reproduction numbers for common infectious dis- eases. It varies from R0 = 1.5 − 1.8, for less contagious diseases like influenza, to R0 = 12 − 18 for the measles and pertussis. Knowing the precise value of R0 is important, but challenging, because of limited testing, inconsistent reporting, and incomplete data [6]. Throughout this book, instead of measuring the basic repro- duction number directly, we estimate it using mathematical modeling and reported case data [17]. Mathematical models interpret the basic reproduction number R0 as the ratio between the infectious period C, the period during which an infectious individual can infect others, and the contact period B, the average time it takes to come into contact with another individual [8], R0 = C/B . (1.1) The longer a person is infectious, and the more contacts the person has during this time, the larger the reproduction number [6]. While we cannot control the infectious period C, we can change our behavior to increase the contact period B [17]. This is precisely what community mitigation strategies and political interventions seek to attempt. The reproduction number of COVID-19. Since the beginning of the coronavirus pandemic, no other number has been discussed more controversially than the re- production number of COVID-19 [22]. The earliest COVID-19 study that followed the first 425 cases of the Wuhan outbreak via direct contact tracing reported a basic reproduction number of R0 = 2.2 [15]. However, especially during the early stages of the outbreak, information was limited because of insufficient testing, changes in case definitions, and overwhelmed healthcare systems. While the concept of R0 seems fairly simple, the reported basic reproduction numbers for COVID-19 vary hugely with country, culture, calculation, and time [17]. Most basic reproduction numbers of COVID-19 we see in the public media today are estimates of mathemat- ical models. These estimates depend critically on the choice of the model, its initial conditions, and many other modeling assumptions [6]. To no surprise, the mathe- matically predicted basic reproduction numbers cover a wide range, from R0 = 2−4 for exponential growth models to R0 = 4 − 7 for more sophisticated compartment models [22].
- 30. 1.4 The basic reproduction number 13 Throughout this book, we identify or infer the reproduction number of COVID- 19–and the uncertainty associated with it–using computational epidemiology [22], Bayesian analysis [10], and reported case data [12]: In the first example for the very early COVID-19 outbreak in China in Section 7.8, we identify a basic reproduction number of R0 = 12.58±3.17 across 30 Chinese provinces. In the second example for the early outbreak in the United States in Section 10.2, we find a basic reproduction number of R0 = 5.30 ± 0.95 across all 54 states and territories. This value suggests that COVID-19 is less contagious than the measles and pertussis with R0 = 12 − 18, as infectious as rubella, smallpox, polio, and mumps with R0 = 4 − 7, slightly more infectious than SARS with R0 = 2 − 5, and more infectious than an influenza with R0 = 1.5 − 1.8. Our basic reproduction number for the United States is significantly lower than our basic reproduction number for China, which could be caused by an increased awareness of COVID-19 transmission a few weeks into the global pandemic. In our third example for the early outbreak in Europe in Section 12.2, the basic reproduction number takes similar values of R0 = 4.62 ± 1.32 across all 27 countries. During these early stages of exponential growth, with new case numbers doubling within two or three days, the most urgent question amongst health care providers and political decision makers was: Can we reduce the reproduction number? For the broad public, this question became famously and illustratively rephrased as: Can we flatten the curve [15]? For the modeling community, the quest for a lower reproduc- tion number all of a sudden meant that traditional epidemiology models were no longer suitable because of changes in the disease dynamics [12]. While traditional models with static parameters were well-suited to model the outbreak dynamics of unconstrained, freely evolving infectious diseases with fixed basic reproduction numbers in the early 20th century, they fail capture how behavioral changes and po- litical interventions can modulate the reproduction number to manage the COVID-19 pandemic in the 21st century [17]. In fact, static reproduction numbers are probably the single most common cause of model failure in COVID-19 modeling [12]. Fortunately, several months into the pandemic, most countries have successfully managed to flatten the new-case-number curves and the reproduction numbers have dropped to values closer to or below one. To model these changes in disease dynamics and reproduction, in Section 7.7, we introduce a dynamic reproduction number, R(t), that accounts for time-varying contact periods. In our fourth example for the European Union in Section 12.2, we show that the initial basic reproduction number of R0 = 4.22±1.69 dropped to an effective reproduction number of R(t) = 0.67±0.18 by mid May 2020. Using machine learning, we correlate mobility and reproduction and identify the responsiveness between the drop in air traffic, driving, walking, and transit mobility and the drop in reproduction to ∆t = 17.24 ± 2.00 days. In the final examples in Sections 13.2 and 13.3 of nine locations across the world, we systematically infer the dynamic reproduction number R(t) throughout a time window of 100 days while accounting for both symptomatic and asymptomatic transmission. From the failure of traditional static epidemiology models [15], we have now learned that we need to introduce dynamic time-varying model parameters if we
- 31. 14 1 Introduction to mathematical epidemiology Fig. 1.6 Basic reproduction numbers and herd immunity thresholds for common infectious diseases. Herd immunity is the indirect protection from an infectious disease that occurs when a large fraction of the population has become immune, either through previous infection or through vaccination. The black line highlights the herd immunity threshold, H = 1 − 1/R0, beyond which this protection occurs, as a function of the basic reproduction number R0. The herd immunity threshold varies from 33-44% for influenza to 92-95% for the measles. want to correctly model behavioral and political changes and reproduce the reported case numbers [17]. This naturally introduces a lot of freedom, a large number of unknowns, and a high level of uncertainty. However, in stark contrast to the epidemic outbreaks in the early 20th century, we now have thoroughly-reported case data and the appropriate tools [23] to address this challenge. The massive amount of COVID- 19 case data, well documented and freely available, has induced a clear paradigm shift from traditional mathematical epidemiology towards data-driven, physics-based modeling of infectious disease [1]. This new technology naturally learns the most probable model parameters–in real time–from the continuously emerging case data, allows us to make projections into the future, and quantifies the uncertainty on the estimated parameters and predictions [26]. 1.5 Concept of herd immunity An important consequence of the basic reproduction number R0 is the condition for herd immunity [9]. Herd immunity describes the indirect protection from an infectious disease that occurs when a large fraction of the population has become immune, either through previous infection or through vaccination [3]. The critical threshold at which the disease reaches this endemic steady state is called the herd immunity threshold, H = 1 − 1/R0 . (1.2) The larger the basic reproduction number R0, the higher the herd immunity threshold H. Table 1.3 and Figure 1.6 summarize the basic reproduction numbers R0 for
- 32. 1.6 Concept of immunization 15 several common infectious diseases along with the estimates of their herd immunity thresholds H. The herd immunity threshold varies from 33-44% for influenza to 92- 95% for the measles. For the reported basic reproduction numbers of R0 = 2.0 − 6.0 of COVID-19, the estimated herd immunity threshold would range from 50-83%. Recent studies that account for the emerging new and more infectious B.1.351 and B.1.1.7 variants of COVID-19 estimate these values to 75-95% [20]. 1.6 Concept of immunization To prevent or revert an epidemic outbreak, we need to ensure that, on average, every infectious individual infects less than one new individual. The concept of herd immunity describes the natural path towards ending an outbreak. However, if the basic reproduction number R0 is large, the herd immunity threshold H = 1 − 1/R0 is high, and waiting for herd immunity through infection alone can be quite devastating. Vaccination is a powerful strategy to accelerate the path towards herd immunity [1]. Fig. 1.7 Vaccination against COVID-19. The objective of COVID-19 vaccination is to provide immunity against severe acute respiratory syndrome coronavirus 2 or SARS-CoV-2, the virus that causes COVID-19 [30]. By May, 2021, thirteen vaccines were authorized for public use: two RNA vaccines (Pfizer–BioNTech and Moderna), five conventional inactivated vaccines (BBIBP-CorV, CoronaVac, Covaxin, WIBP-CorV and CoviVac), four viral vector vaccines (Sputnik V, Oxford–AstraZeneca, Convidecia, and Johnson & Johnson), and two protein subunit vaccines (EpiVacCorona and RBD-Dimer); and more than a billion doses of COVID- 19 vaccines were administered worldwide. Immunization threshold. Vaccination effectively reduces the susceptible population S. Successfully immunizing a fraction I of the population, reduces the susceptible population from S to S∗ 0 = [ 1 − I ] S0 and, with it, the reproduction number from R0 to R∗ 0 = [ 1 − I ] R0. The critical immunization threshold I, below which the effective basic reproduction number is smaller than one, R∗ 0 = [ 1 − I ] R0 < 1, defines the fraction of the population that needs to be immunized to prevent or revert the outbreak of an epidemic, I > 1 − 1/R0 . (1.3) From Table 1.3, we conclude that the required immunization fraction varies signif- icantly, from I > 92 − 95% for the measles with a basic reproduction number of R0 = 12 − 18 to I > 33 − 44% for the common influenza with a basic reproduction number of R0 = 1.5−1.8. This explains, at least in part, why some infectious diseases are a lot more difficult to control through immunization than others.
- 33. 16 1 Introduction to mathematical epidemiology Eradication through vaccination. Once enough individuals are immunized–either through infection or through vaccination–the outbreak stops. A disease that stops circulating in a specific region is considered eliminated in that region. Polio, for example, was eliminated in the United States by 1979 after widespread vaccination efforts. A disease that is eliminated worldwide is considered eradicated. Eradicating a disease through vaccination is a desirable but elusive goal [1]. Malaria has been a candidate for eradication, but although its incidence has been drastically reduced through vaccination, completely eradicating it remains challenging because infection does not result in life long immunity. Polio has been eliminated in most countries through massive vaccination efforts, but still remains present in some regions be- cause its early symptoms often remain unnoticed and infected individuals continue to infect others. Measles have been the target of widespread vaccination, but although the disease is highly recognizable through its characteristic rash, a long latent pe- riod from exposure to the first onset of symptoms complicates outbreak control. To this day, smallpox is the only human infectious disease that has been successfully eradicated through vaccination. The eradication of smallpox is the result of focused surveillance, rapid identification, and ring vaccination [8]. In a massive vaccina- tion campaign launched in 1967, anyone who could have possibly been exposed to smallpox was quickly identified and vaccinated to prevent its further spread. The last known case of smallpox occurred in Somalia in 1977. In 1980 World Health Organization declared smallpox eradicated. The eradication of smallpox remains one of the most notable and profound public health successes in history. Efficacy and risk ratio. The efficacy e of a vaccine is the relative reduction in the disease attack rate between the unvaccinated placebo group npla and the vaccinated group nvac. For a randomized trial with an equal allocation, meaning equally sized placebo and vaccinated groups, npla = nvac, the efficacy is e = npla − nvac npla · 100% = 1 − nvac npla · 100% = 1 − r . (1.4) The risk ratio r is the ratio between the attack rate of the vaccinated group nvac and the placebo group npla, r = nvac npla · 100% = 1 − e . (1.5) The efficacy is an important measure to characterize the success of a vaccine and define critical thresholds below which a vaccination trial should stop. Table 1.4 Contingency table to quantify the significance of a vaccine. The table compares the total number of vaccinated and placebo individuals that developed and did not develop the disease. positive negative total vaccine a b a+b placebo c d c+d total a+c b+d n
- 34. 1.6 Concept of immunization 17 Example: Efficacy of the first COVID-19 vaccine. On November 9, 2020, the Pfizer and BioNTech trial reported a number or COVID-19 cases of ncovid = npla + nvac = 94 and an efficacy of e emin with emin = 90%. The efficacy e of a vaccine is the relative reduction in the disease attack rate between the unvaccinated placebo group npla and the vaccinated group nvac, e = 1 − nvac npla · 100% = 1 − r with r = nvac npla · 100% = 1 − e . where r is the risk ratio. The Pfizer and BioNTech trial was a randomized trial with an equal allocation [30]. Although it did not report detailed numbers, we can estimate the number of placebo and vaccinated cases npla and nvac and the risk ratio r from the reported efficacy emin = 90% with nvac = ncovid − npla, e = 1 − ncovid − npla npla · 100% = 2 − ncovid npla · 100% emin . Solving for the number of placebo cases yields the general equation, npla ncovid 2 − emin , and, for the Pfizer and BioNTech case, npla ncovid/1.1 = 85.45. This implies that for a total of ncovid = 94 COVID-19 positive cases, at a number of placebo cases npla = 86 and vaccinated cases nvac = 8, the efficacy is larger than 90%. Back-calculating the efficacy for these populations, e = 1 − nvac npla · 100% = 1 − 8 86 · 100% = 90.7% 90% = emin , confirms the simulation. According to the protocol, Pfizer and BioNTech planned to take a look at the data at five stages with ncovid = 32,64,92,120,164 reported positive cases and only continue the trial if the efficacy was above e ecrit with ecrit = 62.7%. From this information, we can calculate the critical numbers npla and nvac below which the trial would have stopped at any of the five stages. Solving for the number of placebo cases yields the general equation, npla ncovid 2 − ecrit , and, for the Pfizer and BioNTech case, npla ncovid/1.373. This implies that for ncovid = 32,64,92,120,164, the minimum number of placebo cases to continue the trial was npla ≥ 24,47,68,88,120, for which the resulting efficacies of e = 66.7%,63.8%,64.7%,63.6%,63.3% would all have been slightly above the critical threshold of ecrit = 62.7%.
- 35. 18 1 Introduction to mathematical epidemiology Fisher’s exact test and contingency tables. Fisher’s exact test is a statistical sig- nificance test to analyze contingency tables that quantify the effects of a vaccine compared to a placebo. Table 1.4 illustrates a generic contingency table. From it, we can calculate the case rate across the entire trial as the ratio between the total number of positive cases and the total number of enrolled individuals, νdisease = a + b a + b + c + d with a + b + c + d = ntot . (1.6) The case rates across the vaccinated and placebo groups are, νvac = a a + b with a = nvac and νpla = c c + d with c = npla . (1.7) Most vaccination trials are designed as a randomized trial, meaning they assign participants at random to a vaccinated or placebo group, with equal allocation, meaning they target an equal enrollment into both groups, a + b ≈ c + d ≈ ntot/2. We can use the contingency table to calculate the statistical significance p of the deviation from a null hypothesis, p = a+b a c+d c n a+c = a+b b c+d d n b+d = (a + b)!(c + d)!(a + c)!(b + d)! a! b! c! d! n! , (1.8) in terms of binominal coefficients or factorial operators using a hypergeometric distribution. For a vaccination trial, the lower the p-value, the larger the effect on the vaccinated group compared to the placebo group. Example: Statistical significance of the first COVID-19 vaccine. OnNovem- ber 9, 2020, the Pfizer and BioNTech trial reported ntot = 43,538 enrolled par- ticipants. From this number, we can calculate the case rates across the entire trial, in the vaccinated group, and in the placebo group assuming a random- ization at 1:1 between the vaccinated and placebo groups. The case rate of the entire trial is the ratio between the total number of COVID-19 positive cases and the total number of enrolled individuals, νcovid = ncovid ntot = 94 43,538 = 0.190% . The case rates of the two groups are the ratios between the vaccinated and placebo COVID-19 positive cases and half of the enrolled cases, νvac = nvac ntot/2 = 8 43,538/2 = 0.037% νpla = npla ntot/2 = 86 43,538/2 = 0.395% . Using Fisher’s exact test, we can test the null hypothesis that vaccinated and placebo participants will equally likely contract COVID-19.
- 36. 1.7 Mathematical modeling in epidemiology 19 positive negative total vaccine 8 21,761 21,769 placebo 86 21,683 21,769 total 94 43,444 43,538 We can use the contingency table to estimate the efficiency of the Pfizer BioN- Tech vaccine. Fisher’s exact test calculates the significance of the deviation from a null hypothesis, p = 21769! 21769! 94! 43444! 8! 21761! 86! 21683! 43538! , and confirms, for the Pfizer and BioNTech case with p 0.00001, that indi- viduals in the vaccinated and placebo groups will not equally likely contract COVID-19. 1.7 Mathematical modeling in epidemiology During the early onset of the COVID-19 pandemic, all eyes were on mathematical modeling with the general expectation that mathematical models could precisely predict the trajectory of the pandemic. Mathematical modeling rapidly became front and center to understanding the exponential increase of infections, the shortage of ventilators, and the limited capacity of hospital beds; too rapidly as we now know. Bold and catastrophic predictions not only initiated a massive press coverage, but also a broad anxiety in the general population [15]. However, within only a few weeks, the vastly different predictions and conflicting conclusions began to create the impression that all mathematical models are generally unreliable and inherently wrong [12]. While the failure of COVID-19 modeling–often by an order of mag- nitude and more–was devastating for policymakers and public health practitioners, initial mistakes are not new to the modeling community where an iterative cycle of prediction, failure, and redesign is common standard and best practice [26]. However, the successful use of mathematical models implies to set the expectations right [45]. Understanding what models can and cannot predict is critical to the Art of Mod- eling. Epidemiologists distinguish two kinds of models to understand the outbreak dynamics of an infectious disease: statistical models and mechanistic models [12]. Depending on the degree of complexity, the most popular mechanistic models are compartment models and agent-based models. Statistical models, or more precisely, purely statistical models, use machine learn- ing or regression to analyze massive amounts of data and project the number of infections into the future. The essential idea is to select a function D(t), use statisti- cal tools to fit its coefficients to reported case data D̂(t), and make projections into the future. The function can be quadratic, cubic, logistic, power-law, or exponential,
- 37. 20 1 Introduction to mathematical epidemiology Table 1.5 Mathematical modeling of COVID-19. Summary of the three most common models. Statistical or forecasting models fit nonlinear functions to case data over time, whereas mechanistic models, including compartment and agent-based models, simulate outbreak and contact dynamics. Throughout this book, we use compartment models to simulate the outbreak dynamics of infectious diseases including COVID-19. statistical models mechanistic models forecasting models compartment models agent-based models use machine learning, statis- tics, regression, or method of least squares use physics-based modeling based on nonlinear reaction- diffusion equations use rule-based approaches to study the interaction of au- tonomous systems model case numbers through a nonlinear function model population through compartments model every individual as an independent agent formulate number of cases as a function of time formulate rules by which in- dividuals pass through the compartments formulate simple rules by which individual agents in- teract fit coefficients that are purely phenomenological infer parameters that have a mechanistic interpretation identify parameters that summarize human behavior predicts case numbers from fitting a function predict outbreak dynam- ics, characterize sensitivi- ties, quantify uncertainties predict outbreak dynamics as emergent collective be- havior of individual agents no feedback mechanisms nonlinear feedback discrete contact networks predictions are inexpensive, but very unreliable predictions are reliable only for a small time window predictions are detailed, but computationally expensive for example, D(ϑ,t) = exp(c0 + c1 t + c2 t2 + c3 t3), where D(ϑ,t) are the modeled cumulative cases per day, ϑ = {c0,c1,c2,c3} are the model parameters, and t is the time. One of the simplest statistical tools to compare the model D(ϑ,t) to the data D̂(t) and identify values for its parameters ϑ is the method of least squares. It is important to understand that these parameters are purely phenomenological, they are derived purely by fitting a curve, and typically do not have a mechanistic interpretation. Early in an outbreak, when little is known about disease transmission, epidemiologists often use statistical models because they do not rely on any prior knowledge of the disease. An example for a statistical model of COVID-19 is the ini- tial IHME model [23]. Early in the pandemic, the IHME model used case data from China and Italy to create similar curves, forecast case numbers in the United States, and inform the White House’s response to the pandemic. Carefully constructed sta- tistical frameworks can be used for short-time forecasting using machine learning or regression. This could potentially be useful to understand how to allocate resources or make rapid short-term recommendations. However, purely statistical models can neither capture the dynamics of disease transmission nor the effects of mitigation strategies. This explains, at least in part, why the early COVID-19 predictions based on purely statistical models were off by an order of magnitude or more. To address these serious limitations, several COVID-19 models have now been adjusted to com- bine both statistical modeling and mechanistic modeling. Mechanistic models simulate the outbreak through interacting disease mechanisms by using local nonlinear population dynamics and global mixing of populations
- 38. 1.7 Mathematical modeling in epidemiology 21 [4]. The underlying idea is to identify fundamental mechanisms that drive disease dynamics, for example the duration of the infectious period or the number of con- tacts an infectious individual has during this time. Unlike purely statistical models, mechanistic models include important nonlinear feedback: The more people be- come infected, the faster the disease spreads. By their very nature, the parameters of mechanistic models are not just fitting parameters, they usually have a clear epi- demiological interpretation. This makes mechanistic modeling a powerful strategy to explore different outbreak scenarios or study how an outbreak would change under various assumptions and political interventions [12]. Another advantage of mechanistic models is that we can adjust and improve them dynamically as more information becomes available. Throughout this book, we gradually improve a class of mechanistic models by adding new information. For example, we introduce time- varying dynamic contact rates that vary in different lockdown levels and add the effect of asymptomatic transmission [26]. Even if we do not precisely know the dimension of asymptomatic disease spread, we can use mechanistic models to study what-if scenarios: What would the disease landscape look like if two third of all infectious were asymptomatic? Mechanistic modeling naturally extends into sensi- tivity analysis and uncertainty quantification. As such, it not only provides valuable information about the robustness of the model, but also about the most effective parameters to modulate a disease outbreak [19]. Rather than studying a single one disease trajectory, we could explore a range of trajectories around the mean and characterize the best- and worst-case scenarios. The two most popular mechanistic models in epidemiology are compartment models and agent-based models, and both have been used to understand the outbreak dynamics of COVID-19. When choosing between compartment models and agent-based, it is important to understand the major strengths and weaknesses of each model. Compartment models are the most common approach to model the epidemiology of an infectious disease [14]. Compartment models simulate the collective behavior of subgroups of the population through a number of compartments with labels, for example, SEIR for susceptible, exposed, infectious, and recovered. Individuals move between compartments and the order of the labels indicates the successive motion, for example, SEIS means susceptible, exposed, infectious, then susceptible again [4]. The underlying principle is to model the time evolution of these groups through a set of coupled ordinary differential equations, identify rate constants that characterize their interaction using reported case data, and vary these rate constants to probe different outbreak scenarios. Figure 1.8 illustrates a compartment model that represents the characteristic time- line of COVID-19 through six compartments, the susceptible, exposed, infectious, recovered, hospitalized, and dead groups. This SEIRHD model is defined through a set of six ordinary differential equations that simulate how many individuals reside in each compartment throughout the duration of the outbreak. The model parameters of a compartment model define the transition rates between the individual compart- ments and, for more complex models, the fraction of individuals that transition into a particular path of the disease. For this example, the parameters ϑ = {β,α,γ,νh,νd}
- 39. 22 1 Introduction to mathematical epidemiology Fig. 1.8 Characteristic timeline of COVID-19. On day 0, a fraction of all susceptible individuals is exposed to the virus. After a latent period of A = 1/α = 3 days, the exposed individuals become infectious. After an infectious period of C = 1/γ = 10 days, a fraction (1 − νh) transition to the recovered group, whereas a fraction νh develops severe symptoms and is hospitalized. Of the hospitalized individuals, a fraction (1 − νd) recovers, whereas a fraction νd becomes dead. We can simulate this behavior through an SEIRHD model with six compartments. are the contact rate β, latent rate α, and infectious rate γ, and the hospitalized and dead fractions, νh and νh · νd. From reported case numbers, hospitalizations, and deaths, we can identify or infer the set of model parameters ϑ, the rate constants and fractions, that best explain the model output using statistical tools. Importantly, in contrast to purely statistical models, compartment models are based on model parameters that have a clear physical interpretation. The most important parameters of any compartment model are the contact rate β and the infectious rate γ. Together, they define an important nonlinear feedback that is not present in purely statistical models: The more people become infected, the faster the spread of the disease. The basic reproduction number R0 = β/γ, the ratio of the contact rate β and the infec- tious rate γ, characterizes the magnitude of this feedback. It is an easy-to-understand disease metric that explains how quickly susceptible individuals become infected and how fast a disease spreads across a population [6]. Some epidemiologists ar- gue that, because of their mechanistic nature, compartment models are better suited for long-term predictions than purely statistical models. While this might be true for infectious diseases that develop freely, without any political intervention, the COVID-19 pandemic has taught us that long-term predictions of outbreak dynamics are challenging, even with the most sophisticated mechanistic models [15]. Under- standing the potential and the limitations of compartment modeling is one of the main objectives of this book. In Chapters 2 and 3, we introduce two simple classical compartment models be- fore we introduce the most common compartment model for COVID-19 in Chapter 4. We show how we can use these models to estimate the reproduction number from reported COVID-19 case data. Knowing the precise reproduction number has important consequences for estimating the dimensions of herd immunity and im- munization. Compartment models capture the fundamental dynamics of disease transmission and the effects of public health interventions; however, classical com- partment models ignore the dynamics of the contact rate and its variation across a population. In Section 7.7 we introduce a compartment model that explicitly ac- counts for a time-varying dynamic contact rate, β(t), and captures a varying contact behavior in different subgroups of the population. The dynamic nature of the contact
- 40. 1.8 Data-driven modeling in epidemiology 23 rate naturally introduces a dynamic reproduction number, R(t) = β(t)/γ and allows us to quantify the effectiveness of policy measures as we discuss in Section 12.2. Agent-based models simulate individuals or agents interacting in various social settings and estimate the spread of a disease as these agents come into contact with one another. The underlying idea is to represent each agent individually, formulate relatively simple rules by which individual agents interact, and interpret the collec- tive behavior across all agents as the emergent dynamics of an outbreak. A strength of agent-based models is that they simulate human behavior very granularly: They can assign different parameters or behavior patterns to each individual agent instead of simulating the collective behavior of entire populations. As such, agent-based models offer a lot more freedom than compartment models, but also require a lot more detail. For example, to formulate rules of interaction, agent-based models draw on social connectivity networks, from activity surveys, cell phone locations, public transportation, or airlines statistics. By their very nature, agent-based approaches are computationally expensive, especially for large populations. For small populations, agent-based modeling is a powerful strategy to predict how individual behavior, for example the violation of quarantine, leads to a collective behavior and modulates disease spread. For large populations, agent-based modeling can help rationalize collective model parameters, for example contact rates or reproduction numbers, that feed into more abstract, population level models. Above a certain population size, agent-based models simply become computationally unfeasible and most epi- demiologists would turn to a more macroscopic approach that represents groups of individuals collectively as subgroups of the population. Since our objective is to design and discuss data-driven models for the COVID-19 pandemic, throughout the remainder of this book, we focus exclusively on compartment models. When choosing between purely statistical, compartment, and agent-based models, it is important to know upfront which questions the model should address [45]. Ta- ble 1.5 compares the three models and summarizes their strengths and weaknesses. Throughout this book, we focus on mechanistic compartment modeling. 1.8 Data-driven modeling in epidemiology One year after the World Health Organization had declared the COVID-19 outbreak a global pandemic, SARS-CoV-2 has resulted in more than 118 million reported cases across more than 180 countries and over 2.6 million deaths worldwide. Unlike any other disease in history, the COVID-19 pandemic has generated an unprecedented volume of data, well documented, continuously updated, and broadly available to the general public. There is a critical need for time- and cost-efficient strategies to analyze and interpret these data to systematically manage the pandemic on a global level. Yet, the precise role of physics-based modeling and machine learning in providing quantitative insight into the dynamics of COVID-19 remains a topic of ongoing debate.
- 41. 24 1 Introduction to mathematical epidemiology Physics-based modeling is a successful strategy to integrate multiscale, multi- physics data and uncover mechanisms that explain the dynamics of specific outbreak characteristics. However, physics-based modeling alone often fails to efficiently combine large data sets from different sources and different levels of resolution [26]. Machine learning is as a powerful technique to integrate multimodality, multifidelity data, from cities, counties, states, and countries across the world, and reveal cor- relations between different disease phenomena. However, machine learning alone ignores the fundamental laws of physics and can result in ill-posed problems or non- physical solutions [1]. Throughout this book, we illustrate how data-driven modeling can integrate classical physics-based modeling and machine learning to infer critical disease parameters–in real time–from reported case data to make informed predic- tions and guide political decision making. As a valuable by product, this approach naturally lends itself in sensitivity analysis and uncertainty quantification. From the COVID-19 pandemic, we have learnt that even small inaccuracies in the model can trigger large changes in the number of cases. To understand the vulnerability of the model to these small changes, especially in view of the varying reporting practices of the COVID-19 case data, sensitivity analysis and quantifying uncertainty have become critical elements of robust predictive modeling. Epidemiology data. In data-driven modeling, we need data to fit or infer our model parameters. Unlike earlier pandemics for which case data are often sparse, irregular, or incomplete, the COVID-19 pandemic is amazingly well documented [13]. On hundreds of public COVID-19 dashboards, we can find and download a vast variety of case numbers: daily new cases, active cases, recovered cases, seriously critical cases, cumulative cases, and deaths, at the level of cities, counties, states, countries, or the entire world [9, 12, 25]. Local dashboards often also share the number of hos- pitalizations and intensive care units, which were of great concern especially at the early onset of the pandemic. More recently, these dashboards have also included the number of tests and vaccines. Seroprevalence data with information about the history of the disease, both asymptomatic and symptomatic, are rare and often only avail- able from scientific publications rather than governmental databases. These data are typically updated on a daily basis, and contain notable weekday-weekend alterations. When using the data to infer model parameters and learn about the outbreak behav- ior, we usually smoothen these alteration using seven-day moving averages. Finally, to compare data from different locations, we typically scale the reported case data by the population. A common metric for comparison is the seven-day-per-100,000 incidence, the number of new cases per 100,000 individuals across a seven-day win- dow [38]. Policy makers across the globe use this incidence value to characterize the severity of the outbreak and justify the need for political interventions. Figure 1.9 illustrates a typical data set for the COVID-19 outbreak that we use to infer our epidemiological model parameters. The orange lines summarize the out- break dynamics worldwide throughout the year after the World Health Organization had declared COVID-19 a global pandemic, from March 11, 2020 to March 11, 2021. The light orange lines represent the reported daily new cases, which display notable weekday-weekend fluctuations associated with testing and reporting irreg-
- 42. 1.8 Data-driven modeling in epidemiology 25 Fig. 1.9 Daily new cases of COVID-19 and their seven-day moving average. Daily new COVID- 19 cases are reported on public dashboards worldwide. The light orange curve represents the raw data, the dark orange curve is the seven-day moving average, both reported throughout the year after the WHO had declared COVID-19 a global pandemic on March 11, 2020. ularities. The dark orange lines are the associated seven-day moving average ˆ I(t), which smoothens the fluctuations and displays a much clearer trend in the disease dynamics. Within this one-year window, the absolute daily case numbers peaked on January 8, 2021 with 841,304 new cases. The moving seven-day average peaked on January 11, 2021 with ∆Imax = 745,404 cases. If we assume an infectious pe- riod of C = 1/γ = 7 days, this would result in a maximum infectious population of Imax = 7 · ∆Imax = 7 · 745,404 = 5,217,828, meaning that mid January 2021, more than 5 million people were sick with COVID-19. For a total population of N = 7.8 billion people, this corresponds to a peak seven-day-per-100,000 incidence of Imax · 100,000/N = 67. There are many different ways to use the reported case data. The simplest way is to select a function, use statistical tools to fit its coefficients to the reported case data, and make projections into the future. However, not only the daily new cases, but also the seven-day average in Figure 1.9 display substantial fluc- tuation and it seems difficult to find a function that could explain the orange curves. For data-driven modeling, it is often easier to use the total cumulative case numbers, which always increase monotonically and tend to be more smooth in general. Figure 1.10 shows the total cumulative cases of COVID-19 D̂(t) and a simple statistical model D(t) to fit the data throughout the first year of the pandemic. During the very early stages of an outbreak, an exponential growth model, D(t) = D0 exp(G t), with a growth rate G often provides a good approximation of the total number of cases D̂(t) and is easy to fit. Indeed, this, or similar exponential models, is what many early approaches used. While the initial phase of the outbreak is well represented by exponential growth models, they soon tend to overestimate the outbreak. This is why, during the early COVID-19 pandemic, there was a broad overestimate of the number of cases and of the number of ventilators and hospital beds needed to tread diseased individuals [15]. In this book, instead of using purely statistical models, we use mechanistic models like the compartment model in Figure
- 43. 26 1 Introduction to mathematical epidemiology Fig. 1.10 Total cumulative cases of COVID-19 and a statistical model to fit the data. Cumulative COVID-19 cases, the sum of all daily new cases to that date, are reported on public dashboards worldwide. The solid orange curve represents the raw data, the dashed orange curve is an example of a statistical model, both reported throughout the year after the WHO had declared COVID-19 a global pandemic on March 11, 2020. 1.8. For this SEIRHD compartment model, the number of total cases is the sum of all infectious, hospitalized, recovered and dead individuals, I(t) + R(t) + H(t) + D(t). Throughout this book, we use different types of compartment models, infer their model parameters using reported COVID-19 case data from public databases [9, 12, 25], seroprevalence data from scientific publications [26], and mobility data [2, 10, 10], and vary the parameters to probe different outbreak scenarios. Problems 1.1 Testing, testing, testing. On June 6, 2020, Donald J. Trump, the President of the United States, famously said “Remember this, when you test more, you have more cases.” What is wrong with this statement? What did he really mean to say? Discuss the implications of the testing frequency on reported case numbers, and, ultimately, on policy making. 1.2 Herd immunity. On September 15, 2020, Donald J. Trump, the President of the United States, claimed that COVID-19 would go away, without a vaccine. “You’ll develop herd. Like a herd mentality.” What did he really mean to say? Discuss the effects of vaccination on herd immunity. 1.3 Herd immunity. Assume the basic reproduction number of the swine flu, caused by the H1N1 virus, was on the order of R0 = 1.4 − 1.6. Calculate its herd immunity threshold H and compare it against the herd immunity thresholds of other infectious diseases in Table 1.3 and Figure 1.6. Comment on whether this makes the swine flu a good candidate for eradication through vaccination.
- 44. 1.8 Data-driven modeling in epidemiology 27 1.4 COVID-19 variants B.1.351 and B.1.1.7. Assume the basic reproduction num- ber of COVID-19 in December 2020 was R0 = 2.0. Calculate the herd immunity threshold H. Now, assume that the variant B.1.351 with a 50% increased basic re- production number has been introduced into the population [20]. How does the herd immunity threshold change? How would the variant B.1.1.7 with a 56% increased basic reproduction number change the herd immunity threshold? Comment on your results. Fig. 1.11 The Lessons of the Pandemic. Twelve condensed rules to the the avoidance of unneces- sary personal risks and to the promotion of better personal health by George A. Soper in reflection of the Spanish flu in 1919. 1.5 The Lessons of the Pandemic. In his Science publication The Lessons of the Pandemic, George A. Soper lists twelve public health measures to manage the Spanish flu in 1919. Read the twelve recommendations in Figure 1.11 and comment on which of them are still valid for the COVID-19 pandemic today. 1.6 Efficacy of the AstraZeneca COVID-19 vaccine. The randomized, equal allo- cation AstraZeneca trial enrolled 11636 participants in its interim efficacy analysis, 5807 in the vaccinated group and 5829 in the unvaccinated placebo group. Of the vaccinated group, 30 developed COVID-19, and of the placebo group 101. Calculate the efficacy and risk ratio of the AstraZeneca vaccine. 1.7 Case rates during the AstraZeneca COVID-19 trial. The AstraZeneca trial en- rolled 11636 participants in its interim efficacy analysis, 5807 in the vaccinated group and 5829 in the unvaccinated placebo group. Of the vaccinated group, 30 developed COVID-19, and of the placebo group 101. Create a contingency table for the interim efficacy analysis. Calculate the overall case rate of the trial and the case rates in the vaccinated and placebo groups. 1.8 Efficacy of the Janssen COVID-19 vaccine. The randomized, equal allocation Janssen COVID-19 trial enrolled 39,321 participants, 19,630 received the vaccine
- 45. 28 1 Introduction to mathematical epidemiology and 19,691 received placebo. At least 14 days after vaccination, the trial recorded 116 COVID-19 cases in the vaccine group and 348 cases in the placebo group. Calculate the efficacy of the Janssen COVID-19 vaccine and compare it to the efficacy of the Pfizer BioNTech and AstraZeneca vaccines. How would you expect the efficacy to change after another 14 days? 1.9 Efficacy of the Janssen COVID-19 vaccine. The randomized, equal allocation Janssen COVID-19 trial enrolled 39,321 participants, 19,630 received the vaccine and 19,691 received placebo. At least 28 days after vaccination, the trial recorded 66 COVID-19 cases in the vaccine group and 193 cases in the placebo group. Calculate the efficacy of the Janssen COVID-19 vaccine after 28 days and compare it to the efficacy after 14 days. Comment on why you would or would not have expected this result. 1.10 Statistical models. Find and download the total cumulative COVID-19 case data D̂(t) from your own city, county, state, country, or the world. Plot the case data similar to Figure 1.10. Try to fit a linear function D(t) = D0 + c1 t and a quadratic function D(t) = D0 + c1 t + c2 t2 to the very early stages of the outbreak. Which function is easier to fit? What does that tell you about the early outbreak? 1.11 Statistical models. Find and download the total cumulative COVID-19 case data D̂(t) from your own city, county, state, country, or the world. Plot the case data similar to Figure 1.10. Fit an exponential function D(t) = D0 exp(G t) to the early stages of the outbreak. What is your growth rate G? When does the exponential function fail to describe the case data D̂(t)? 1.12 Epidemiology data. Find and download the daily new COVID-19 case data from your own city, county, state, country, or the world. Plot the raw data similar to Figure 1.9. Explain the local fluctuations in the reported case data that occur on the order of days. 1.13 Epidemiology data. Find and download the daily new COVID-19 case data from your own city, county, state, country, or the world. Calculate and plot the seven- day moving average similar to Figure 1.9. How many waves can you identify? Explain the global fluctuations that occur on the order of weeks or months. Interpret the growth or decay in case numbers in view of specific events or political interventions. 1.14 Epidemiology data. Find the COVID-19 case data from your own state or country and your furthest away vacation destination. Compare the different outbreak dynamics. Do you think the reporting between both locations is consistent? Identify at least four potential sources of error in reporting daily COVID-19 case data. 1.15 Incidence. Find and download the daily new COVID-19 case data from your own city, county, state, country, or the world. Calculate the seven-day moving average. Identify the peak seven-day moving average ∆Imax within your simulation window. Assume an infectious period of C = 1/γ = 7 days. Calculate the maximum infectious population Imax = 7·∆Imax. Find the total population N of your location and calculate its peak seven-day-per-100,000 incidence, Imax · 100,000/N.
- 46. References 29 1.16 Incidence and the effect of scale. Studying an outbreak at a more global scale tends to smoothen fluctuations and local peaks. Find and download the daily new COVID-19 case data from the next smaller or larger scale compared to the previous problem. If you have studied your state, now study your city, county, or country. Identify the peak seven-day moving average ∆Imax within your simulation window and calculate the maximum infectious population Imax = 7 · ∆Imax. Find the total population N of your location and calculate its peak seven-day-per-100,000 incidence, Imax ·100,000/N. Compare you results against the seven-day-per-100,000 incidence at the smaller or larger scale. Interpret your results. 1.17 Epidemiology data and compartment models. The compartment model in Figure 1.8 introduces three possible disease paths from infection: direct recovery at a fraction (1 − νh), hospitalization and recovery at a fraction νh (1 − νd), and hospitalization and death at a fraction νh · νd. The fraction νh · νd is called the case fatality rate and is about 2% worldwide for COVID-19. Find extreme values for the case fatality rate. Discuss which factors influence the case fatality rate both globally and locally. 1.18 Epidemiology data and compartment models. Assume you want to learn parameters for your compartment model, for example the one in Figure 1.8, from reported case data, hospitalizations, and deaths. Early in the pandemic, when testing was slow and reporting was delayed, epidemiologists suggested to use deaths rather than daily new cases for model calibration. This initiated a controversial and still ongoing discussion how to count COVID-19 deaths. Discuss the difference between death from and death with COVID-19 in terms of absolute numbers, case fatality ratios, and model parameters. References 1. Alber M, Buganza Tepole A, Cannon W, De S, Dura-Bernal S, Garikipati K, Karniadakis G, Lytton WW, Perdikaris P, Petzold L, Kuhl E (2019) Integrating machine learning and multiscale modeling: Perspectives, challenges, and opportunities in the biological, biomedical, and behavioral sciences. npj Digital Medicine 2:115. 2. Anderson RM, May RM (1982) Directly transmitted infectious diseases: control by vaccination. Science 215:1053-1060. 3. Anderson RM, May RM (1991) Infectious Diseases of Humans. Oxford University Press, Oxford. 4. Apple Mobility Trends. https://www.apple.com/covid19/mobility. accessed: June 1, 2021. 5. Bernoulli D (1760) Essay d’une nouvelle analyse de la mortalite causee par la petite verole et des avantages de l’inoculation pour la prevenir. Mémoires de Mathématiques et de Physique, Académie Royale des Sciences, Paris 1-45. 6. Brauer F, Castillo-Chavez C (2001) Mathematical Models in Population Biology and Epidemi- ology. Springer-Verlag New York. 7. Brauer F, van den Dreissche P, Wu J (2008) Mathematical Epidemiology. Springer-Verlag Berlin Heidelberg.
- 47. 30 1 Introduction to mathematical epidemiology 8. Brauer F (2017) Mathematical epidemiology: Past, present and future. Infectious Disease Modelling 2:113-127. 9. Brauer F, Castillo-Chavez C, Feng Z (2019) Mathematical Models in Epidemiology. Springer- Verlag New York. 10. Delamater PL, Street EJ, Leslie TF, Yang YT, Jacobsen KH (2019) Complexity of the basic reproduction number (R0). Emerging Infectious Diseases 25:1-4. 11. Dieckmann O, Heesterbeek JAP (2000) Mathematical Epidemiology of Infectious Diseases: Model Building, Analysis and Interpretation. Wiley. 12. Dietz K (1993) The estimation of the basic reproduction number for infectious diseases. Statistical Methods in Medical Research 2:23-41. 13. Dong E, Gardner L (2020) An interactive web-based dashboard to track COVID-19 in real time. The Lancet Infectious Diseases 20:533-534. 14. European Centre for Disease Prevention and Control. Situation update worldwide. https: //www.ecdc.europa.eu/en/geographical-distribution-2019-ncov-cases accessed: June 1, 2021. 15. Eurostat. Your key to European statistics. Air transport of passengers. https://ec.europa. eu/eurostat accessed: June 1, 2021. 16. Evans AS (1976) Viral Infections of Humans. Epidemiology and Control. Plenum Medical Book Company, New York and London. 17. Fauci AS, Lane HC, Redfield RR (2020) Covid-19–Navigating the uncharted. New England Journal of Medicine 382:1268-1269. 18. Fine PEM (1993) Herd immunity: history, theory, practice. Epidemiologic Reviews 15:265- 302. 19. Gelman A, Carlin JB, Stern HS, Dunson DB, Vektari A, Rubin DB (2013) Bayesian Data Analysis. Chapman and Hall/CRC, 3rd edition. 20. Gorbalenya AE, Baker SC, Baric RS, de Groot RJ, Drosten C, Gulyaeva AA, Haagmans BL, Lauber C, Leontovich AM, Neuman BW, Penzar D, Perlman S, Poon LLM, Samborskiy D, Sidorov IA, Sola I, Ziebuhr J (2020) Severe acute respiratory syndrome-related coronavirus: the species and its viruses-a statement of the coronavirus study group. Nature Microbiology 5:536-544. 21. Hethcote HW (2000) The mathematics of infectious diseases. SIAM Review 42:599-653. 22. Holmdahl I, Buckee C. Wrong but useful–What Covid-19 epidemiolgic models can and cannot tell us. New England Journal of Medicine 383:303-305. 23. Institute for Health and Metrics Evaluation IHME. COVID-19 Projections. https://covid1 9.healthdata.org. assessed: July 27, 2020. 24. International Air Transport Association (2020) https://www.iata.org. accessed: July 9, 2020. 25. Ioannidis JPA, Cripps S, Tanner MA (2021) Forecasting for COVID-19 has failed. International Journal of Forecasting, in press. 26. Johns Hopkins University (2021) Coronavirus COVID-19 Global Cases by the Center for Systems Science and Engineering. https://coronavirus.jhu.edu/map.html, https: //github.com/CSSEGISandData/covid-19 assessed: June 1, 2021. 27. Kermack WO, McKendrick G (1927) Contributions to the mathematical theory of epidemics, Part I. Proceedings of the Royal Society London Series A 115:700-721. 28. Krämer A, Kretzschmar M, Krickeberg K (2010) Modern Infectious Disease Epidemiology. Springer-Verlag New York. 29. Kuhl E (2020) Data-driven modeling of COVID-19 – Lessons learned. Extreme Mechanics Letters 40:100921. 30. Kyriakidis NC, Lopez-Cortes A, Vasconez Gonzalez E, Barreto Grimaldos A, Ortiz Prado E. SARS-CoV-2 vaccines strategies: a comprehensive review of phase 3 candidates. npj Vaccines 6:28. 31. Li Q, Guan X, Wu P, Wang X, ... Feng Z (2020) Early transmission dynamics in Wuhan, China, of novel coronavirus-infected pneumonia. New England Journal of Medicine 382:1199-1207. 32. Liang ST, Liang LT, Rosen JM (2021) COVID-19: a comparison to the 1918 influenza and how we can defeat it. BMJ Postgraduate Medical Journal 97:273-274
- 48. References 31 33. Linka K, Peirlinck M, Kuhl E (2020) The reproduction number of COVID-19 and its correlation with public heath interventions. Computational Mechanics 66:1035-1050. 34. Linka K, Goriely A, Kuhl E (2021) Global and local mobility as a barometer for COVID-19 dynamics. Biomechanics and Modeling in Mechanobiology 20:651–669. 35. Linka K, Peirlinck M, Schafer A, Ziya Tikenogullari O, Goriely A, Kuhl E (2021) Effects of B.1.1.7 and B.1.351 on COVID-19 dynamics. A campus reopening study. Archives of Computational Methods in Engineering. doi:10.1007/s11831-021-09638-y. 36. Liu J, Shang, X (2020) Computational Epidemiology. Springer International Publishing. 37. Liu Y, Gayle AA, Wilder-Smith A, Rocklöv J (2020) The reproductive number of COVID-19 is higher compared to SARS coronavirus. Journal of Travel Medicine (2020) 27:taaa021. 38. Lu H, Weintz C, Pace J, Indana D, Linka K, Kuhl E (2021) Are college campuses superspread- ers? A data-driven modeling study. Computer Methods in Biomechanics and Biomedical Engineering doi:10.1080/10255842.2020.1869221. 39. Martcheva M (2015) An Introduction to Mathematical Epidemiology. Springer Science + Business Media New York. 40. New York Times (2020) Coronavirus COVID-19 Data in the United States. https://github .com/nytimes/covid-19-data/blob/master/us-states.csv assessed: June 1, 2021. 41. Osvaldo M (2018) Bayesian Analysis with Python: Introduction to Statistical Modeling and Probabilistic Programming Using PyMC3 and ArviZ. Packt Publishing, 2nd edition. 42. Paul JR (1966) Clinical Epidemiology. University of Chicago Press, Chicago. 43. Peirlinck M, Linka K, Sahli Costabal F, Bendavid E, Bhattacharya J, Ioannidis J, Kuhl E (2020) Visualizing the invisible: The effect of asymptomatic transmission on the outbreak dynamics of COVID-19. Computer Methods in Applied Mechanics and Engineering 372:113410. 44. Peng GCY, Alber M, Buganza Tepole A, Cannon W, De S, Dura-Bernal S, Garikipati K, Karniadakis G, Lytton WW, Perdikaris P, Petzold L, Kuhl E (2021) Multiscale modeling meets machine learning: What can we learn? Archive of Computational Methods in Engineering 28:1017-1037. 45. Siegenfeld AF, Taleb NN, Bar-Yam Y (2020) What models can and cannot tell us about COVID-19. Proceedings of the National Academy of Sciences 117:16092-16095. 46. Snow J (1855) On the Mode of Communication of Cholera (2nd edition). London, John Churchill. 47. Soper GA (1919) The lessons of the pandemic. Science XLIX 501-506. 48. Viceconte G, Petrosillo N (2020) COVID-19 R0: Magic number or conundrum? Infectious Disease Reports 12:8516. 49. World Health Organization. WHO Virtual Press Conference on COVID-19, March 11, 2020 https://www.who.int/docs/default-source/coronaviruse/transcripts/who-a udio-emergencies-coronavirus-press-conference-full-and-final-11mar2020 .pdf?sfvrsn=cb432bb32 accessed: June 1, 2021.
- 49. Chapter 2 The classical SIS model Abstract The SIS model is the simplest compartment model with only two pop- ulations, the susceptible and infectious groups S and I. It characterizes infectious diseases like the common cold or influenza that do not provide immunity upon infection. While the SIS model is too simplistic to explain the outbreak dynamics of complex infectious diseases, it is the only compartment model with an explicit analytical solution for the time course of its populations. This makes it a popular model to explain and illustrate the basic principles of compartment modeling. The learning objectives of this chapter on classical SIS modeling are to • explain concepts of mass action incidence and constant rate recovery • interpret contact and infectious rates and periods • solve the analytical solutions for the susceptible and infectious populations • distinguish disease free and endemic equilibria • analyze the final size relation • demonstrate how the infectious rate, reproduction number, and initial conditions modulate outbreak dynamics • discuss limitations of classical SIS modeling By the end of the chapter, you will be able to analyze, simulate, and predict the outbreak dynamics of simple infectious diseases like the common cold or influenza that do not provide immunity to reinfection. 2.1 Introduction of the SIS model Some infectious diseases, for example from the common cold or influenza, do not provide lifelong immunity upon recovery from infection [2, 2], and previously in- fected individuals become susceptible again [10]. We can model their behavior through the simplest of all compartment models, the SIS model [5]. Figure 11.4 shows that the SIS model consists of only two populations, the susceptible group S and the infectious group I [3]. As such, it provides valuable insight into the concept 33 © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 E. Kuhl, Computational Epidemiology, https://doi.org/10.1007/978-3-030-82890-5_2
- 50. 34 2 The classical SIS model Fig. 2.1 Classical SIS model. The classical SIS model contains two compartments for the sus- ceptible and infectious populations, S and I. The transition rates between the compartments, the contact and infectious rates, β and γ, are inverses of the contact and infectious periods, B = 1/β and C = 1/γ. of compartment modeling [8]. The SIS model is often used to illustrate the basic features of compartment models because we can solve the dynamics of its two pop- ulations, S(t) and I(t), analytically at any point t in time [7]. The SIS model also has simple analytical solutions for the converged final sizes, S∞ and I∞, as t → ∞ [13]. For the transition from the susceptible to the infectious group, we assume a mass action incidence [4], which implies that the rate of new infections is proportional to the size of the susceptible and infectious groups S and I weighted by the contact rate β, Û I = β SI. For the transition from the infectious to the susceptible group, we Fig. 2.2 The transition from the susceptible to the infectious group is based on the assumption of mass action incidence for which the rate of new infections is proportional to the size of the susceptible and infectious groups S and I, weighted by the contact rate β, Û I = β SI. The transition from the infectious to the susceptible group is based on the assumption of constant rate recovery for which the rate of recovered infections is proportional to the size of the infectious group I, weighted by the infectious rate γ, Û I = −γI. assume a constant rate recovery, which implies that the rate of recovered infections is proportional to the size of the infectious group I weighted by the infectious rate γ, Û I = γI. Figure 2.2 illustrates the dynamics of these two assumptions that result in the following system of two coupled ordinary differential equations, Û S = − β SI + γ I Û I = + β SI − γ I . (2.1) The transition rates between both compartments are the contact rate β and the infectious rate γ in units [1/days], which are the inverses of the contact period B = 1/β and the infectious period C = 1/γ in units [days]. The ratio between the contact and infectious rates, or similarly, between the infectious and contact periods, defines the basic reproduction number R0, R0 = β γ = C B . (2.2) In this simple format, the SIS model (2.1) neglects all vital dynamics, Û S + Û I 0 and S + I = const. = 1. It does not account for births, natural deaths, or death from the disease.
- 51. 2.2 Analytical solution of the SIS model 35 2.2 Analytical solution of the SIS model From the condition of non-vital dynamics, S = 1 − I, we can rephrase the system of equations of the SIS model (2.1) in terms of only one independent variable, the size of the infectious group I, governed by the following nonlinear ordinary differential equation, Û I = β [ 1 − I ] I − γ I = [ β − γ ] I 1 − I 1 − γ/β . (2.3) Equation (2.3) is a logistic differential equation of the form, Û I = r I [ 1 − I/K ] with r = β − γ and K = 1 − γ/β, (2.4) where K is the carrying capacity. This type of equation has an explicit analytical solution, I(t) = K I0 I0 + [ K − I0 ] exp (−r t) , (2.5) where I0 = I(0) is the initial infectious population [10]. It proves convenient to reparameterize the equation for the infectious population (2.5) in terms of the basic reproduction number R0 = β/γ and the infectious period C = 1/γ with r = [R0 − 1]/C and K = 1 − 1/R0. This provides the analytical solution for the SIS model in terms of the infectious period C, the basic reproduction number R0, and the initial infectious population I0, S(t) = 1− [1 − 1/R0] I0 I0 + [1 − 1/R0 − I0] exp ([1 − R0] t/C) I(t) = [1 − 1/R0] I0 I0 + [1 − 1/R0 − I0] exp ([1 − R0] t/C) . (2.6) Figures 2.3, 2.4, and 2.5 highlight the outbreak dynamics of the SIS model, solved analytically using equations (2.6), for the time period of one year. The three figures demonstrate the sensitivity of the SIS model for varying infectious periods C, and varying basic reproduction numbers R0, and initial infectious populations I0. Increas- ing the infectious period C delays convergence to the endemic equilibrium, but the final sizes S∞ and I∞ remain unchanged. Increasing the basic reproduction number R0 accelerates convergence to the endemic equilibrium, decreases S∞, and increases I∞. Increasing the initial infectious population I0 accelerates the onset of the out- break, but the final sizes S∞ and I∞ remain unchanged. Interestingly, increasing the initial exposed population I0 by an order of magnitude shifts the population dynam- ics by a constant time increment, for the current parameterization by 50 days. This highlights the exponential nature of the SIS model, which causes a constant acceler- ation of the outbreak for a logarithmic increase of the initial infectious population, while the overall outbreak dynamics remain the same.
- 52. 36 2 The classical SIS model Fig. 2.3 Classical SIS model. Sensitivity with respect to the infectious period C. Increasing the infectious period C delays convergence to the endemic equilibrium, but the final sizes S∞ and I∞ remain unchanged. Basic reproduction number R0 = 2.0, initial infectious population I0 = 0.01, and infectious period C = 5, 10, 15, 20, 25, 30 days. Fig. 2.4 Classical SIS model. Sensitivity with respect to the basic reproduction number R0. Increasing the basic reproduction number R0 accelerates convergence to the endemic equilibrium, decreases S∞, and increases I∞. Infectious period C = 20 days, initial infectious population I0 = 0.01, and basic reproduction number R0 = 1.5, 1.7, 2.0, 2.4, 3.0, 5.0, 10.0. 2.3 Final size relation of the SIS model For practical purposes, it is interesting to estimate the final susceptible and infectious populations S∞ and I∞ [15]. From the analytical solution (2.5), I∞ = K I0 I0 + [ K − I0 ] exp (−r t∞) with r = β − γ and K = 1 − γ/β, (2.7)
- 53. 2.3 Final size relation of the SIS model 37 Fig. 2.5 Classical SIS model. Sensitivity with respect to the initial infectious population I0. Increasing the initial infectious population I0 accelerates the onset of the outbreak, but the final sizes S∞ and I∞ remain unchanged. Infectious period C = 20 days, basic reproduction number R0 = 2.0, and initial infectious population I0 = 10−1, 10−2, 10−3, 10−4, 10−5, 10−6. it is easy to show that we can distinguish two cases. If there is a finite initial infectious population, I0 0, the infectious population will converge to either zero, for β γ, r 0, and exp (−r t∞) → ∞, or to K for β γ, r 0, and exp (−r t∞) → 0 [2, 9], thus I∞ = 0 if β γ K if β γ . (2.8) We can rephrase these limit conditions (2.8) in terms of the basic reproduction number R0. In classical epidemiology, the two converged states are known as the disease-free equilibrium and the endemic equilibrium [3]. The equations that define the final converged populations of the endemic equilibrium state are the final size relation. R0 1 disease-free equilibrium: S∞ = 1 and I∞ = 0 R0 1 endemic equilibrium: S∞ = 1/R0 and I∞ = 1 − 1/R0 . (2.9) Figure 2.6 illustrates the final size relation and emphasizes the role of the basic reproduction number R0 as the distinguishing outbreak characteristic between the disease-free and endemic equilibrium. For reproduction numbers smaller than one, R0 1, the SIS model converges to a disease-free equilibrium with S∞ = 1 and I∞ = 0. For reproduction numbers larger than one, R0 1, the SIS model converges to an endemic equilibrium with S∞ = 1/R0 and I∞ = 1 − 1/R0. The larger the reproduction number R0, the smaller the final susceptible population S∞ and the larger the final infectious population I∞ [3].
- 54. 38 2 The classical SIS model Fig. 2.6 Classical SIS model. Final size relation as a function of the basic reproduction num- ber R0. For R0 1, the SIS model converges to a disease-free equilibrium with S∞ = 1 and I∞ = 0. For R0 1, the SIS model converges to an endemic equilibrium with S∞ = 1/R0 and I∞ = 1 − 1/R0. Increasing R0 reduces the susceptible population S∞ and increases the infectious population I∞. Problems 2.1 Basic reproduction number. Throughout the winter of 1980, a Scottish board- ing school reported an influenza that, on any day, affected on average 408 of its 1632 students. Estimate the basic reproduction number. 2.2 Basic reproduction number. Throughout the winter of 1980, a Scottish board- ing school reported an influenza with an infectious period of C = 4 days and a contact period of B = 2.8 days. How many of its 1632 students are infectious at endemic equilibrium? 2.3 Basic reproduction number. Assume the common flu has an infectious period of C = 7 days and a contact rate of β = 0.2/ days. Determine the basic reproduction number R0 and the final sizes of the susceptible and infectious populations S∞ and I∞. 2.4 Contact rate. Assume the common flu has an infectious period of C = 7 days. Determine the contact rate β for which the infectious population I never increases beyond 20% of the population. What is the basic reproduction number under these conditions? 2.5 Contact rate. Assume the common flu has an infectious period of C = 7 days. Determine the contact rate β for which the infectious population I never increases beyond 30% of the population. Comment on how the basic reproduction number for a maximum infectious population of 30% differs to the basic reproduction number for a maximum infectious population of 20%.
- 55. Random documents with unrelated content Scribd suggests to you:
- 56. the cords used in tying the blankets into bales, which same cross lines appear as cords in l, Fig. 33. Mr. Coronel also possesses small figures of Mexicans, of various conditions of life, costumes, trades, and professions, one of which, a painted statuette, is a representation of a Mexican lying down flat upon an outspread serape, similar in color and form to the black and white bands shown in the upper figure of d, Fig. 32, and a, b, of Fig. 33, and instantly suggesting the explanation of those figures. Upon the latter the continuity of the black and white bands is broken, as the human figures are probably intended to be in front, or on top, of the drawings of the blankets. Fig. 33.—Petroglyph in Santa Barbara county, California. The small statuette above mentioned is that of a Mexican trader, and if the circles in the petroglyphs are considered to represent bales of blankets, the character in Fig. 32, d, is still more interesting, from the union of one of these circles with a character representing the trader, i. e., the man possessing the bales. Bales, or what appear to be bales, are represented to the top and right of the circle in d, in that figure. In Fig. 33, l, a bale is upon the back of what appears to be a horse, led in an upward direction by an Indian whose headdress and ends of the breechcloth are visible. To the right of the bale are three short lines, evidently showing the knot or ends of the
- 57. cords used in tying a bale of blankets without colors, therefore of less importance, or of other goods. Other human forms appear in the attitude of making gestures, one also in j, Fig. 33, probably carrying a bale of goods. In the same figure u represents a centipede, an insect found occasionally south of the mountains, but reported as extremely rare in the immediate northern regions. For remarks upon x in the same figure see Chapter xx, Section 2, under the heading The Cross. Mr. Coronel stated that when he first settled in Los Angeles, in 1843, the Indians living north of the San Fernando mountains manufactured blankets of the fur and hair of animals, showing transverse bands of black and white similar to those depicted, which were sold to the inhabitants of the valley of Los Angeles and to Indians who transported them to other tribes. It is probable that the pictographs are intended to represent the salient features of a trading expedition from the north. The ceiling of the cavity found between the paintings represented in the two figures has disappeared, owing to disintegration, thus leaving a blank about 4 feet long, and 6 feet from the top to the bottom between the paintings as now presented. COLORADO. Petroglyphs are reported by Mr. Cyrus F. Newcomb as found upon cliffs on Rock creek, 15 miles from Rio Del Norte, Colorado. Three small photographs, submitted with this statement, indicate the characters to have been pecked; they consist of men on horseback, cross-shaped human figures, animals, and other designs greatly resembling those found in the country of the Shoshonean tribes, examples of which are given infra. Another notice of the same general locality is made by Capt. E. L. Berthoud (a) as follows:
- 58. The place is 20 miles southeast of Rio Del Norte, at the entrance of the canyon of the Piedra Pintada (Painted rock) creek. The carvings are found on the right of the canyon or valley and upon volcanic rocks. They bear the marks of age and are cut in, not painted, as is still done by the Utes everywhere. They are found for a quarter of a mile along the north wall of the canyon, on the ranches of W. M. Maguire and F. T. Hudson, and consist of all manner of pictures, symbols, and hieroglyphics done by artists whose memory even tradition does not now preserve. The fact that these are carvings done upon such hard rock invests them with additional interest, as they are quite distinct from the carvings I saw in New Mexico and Arizona on soft sandstone. Though some of them are evidently of much greater antiquity than others, yet all are ancient, the Utes admitting them to have been old when their fathers conquered the country. Mr. Charles D. Wright, of Durango, Colorado, in a communication dated February 20, 1885, gives an account of some “hieroglyphs” on rocks and upon the walls of cliff houses near the boundary line between Colorado and New Mexico. He says: The following were painted in red and black paints on the wall (apparently the natural rock wall) of a cliff house: At the head was a chief on his horse, armed with spear and lance and wearing a pointed hat and robe; behind this character were some twenty characters representing people on horses lassoing horses, etc. In fact the whole scene represented breaking camp and leaving in a hurry. The whole painting measured about 12 by 16 feet. Mr. Wright further reports characters on rocks near the San Juan river. Four characters represent men as if in the act of taking an obligation, hands extended, and wearing a “kind of monogram on breast, and at their right are some hieroglyphics written in black paint covering a space 3 by 4 feet.”
- 59. The best discussed and probably the most interesting of the petroglyphs in the region are described and illustrated by Mr. W. H. Holmes (a), of the Bureau of Ethnology. The illustrations are here reproduced in Figs. 34 to 37, and the remarks of Mr. Holmes, slightly condensed, are as follows: The forms reproduced in Fig. 34 occur on the Rio Mancos, near the group of cliff houses. They are chipped into the rock evidently by some very hard implement and rudely represent the human figure. They are certainly not attempts to represent nature, but have the appearance rather of arbitrary forms, designed to symbolize some imaginary being. Fig. 34.—Petroglyphs on the Rio Mancos, Colorado. The forms shown in Fig. 35 were found in the same locality, not engraved, but painted in red and white clay upon the smooth rocks. These were certainly done by the cliff-builders, and probably while the houses were in process of construction, since the material used is identical with the plaster of the houses. The sketches and notes were made by Mr. Brandegee. The reproduction is approximately one-twelfth the size of the original.
- 60. Fig. 35.—Petroglyphs on the Rio Mancos, Colorado. The examples shown in Fig. 36 occur on the Rio San Juan about 10 miles below the mouth of the Rio La Plata and are actually in New Mexico. A low line of bluffs, composed of light-colored massive sandstones that break down in great smooth-faced blocks, rises from the river level and sweeps around toward the north. Each of these great blocks has offered a very tempting tablet to the graver of the primitive artist, and many of them contain curious and interesting inscriptions. Drawings were made of such of these as the limited time at my disposal would permit. They are all engraved or cut into the face of the rock, and the whole body of each figure has generally been chipped out, frequently to the depth of one-fourth or one-half of an inch.
- 61. Fig. 36.—Petroglyphs on the Rio San Juan, New Mexico. The work on some of the larger groups has been one of immense labor, and must owe its completion to strong and enduring motives. With a very few exceptions the engraving bears undoubted evidence of age. Such new figures as occur are quite easily distinguished both by the freshness of the chipped surfaces and by the designs themselves. The curious designs given in the final group have a very perceptible resemblance to many of the figures used in the embellishment of pottery. The most striking group observed is given in Fig. 37 a, same locality. It consists of a great procession of men, birds, beasts, and fanciful figures. The whole picture as placed upon a rock is highly spirited and the idea of a
- 62. general movement toward the right, skillfully portrayed. A pair of winged figures hover about the train as if to watch, or direct its movements; behind these are a number of odd figures, followed by an antlered animal resembling a deer, which seems to be drawing a notched sledge containing two figures of men. The figures forming the main body of the procession appear to be tied together in a continuous line, and in form resemble one living creature about as little as another. Many of the smaller figures above and below are certainly intended to represent dogs, while a number of men are stationed about here and there as if to keep the procession in order. Fig. 37.—Petroglyphs on the Rio San Juan, New Mexico. As to the importance of the event recorded in this picture, no conclusions can be drawn; it may represent the migration of a tribe or family or the trophies of a victory. A number of figures are wanting in the drawing at the left, while some of those at the right may not belong properly to the main group. The reduction is, approximately, to one-twelfth. Designs B and C of the same figure represent only the more distinct portions of two other groups. The
- 63. complication of figures is so great that a number of hours would have been necessary for their delineation, and an attempt to analyze them here would be fruitless. It will be noticed that the last two petroglyphs are in New Mexico, but they are so near the border of Colorado and so connected with the series in that state that they are presented under the same heading. CONNECTICUT. The following account is extracted from Rafn’s Antiquitates Americanæ (a): In the year 1789 Doctor Ezra Stiles, D. D., visited a rock situated in the Township of Kent in the State of Connecticut, at a place called Scaticook, by the Indians. He thus describes it: “Over against Scaticook and about one hundred rods East of Housatonic River, is an eminence or elevation which is called Cobble Hill. On the top of this stands the rock charged with antique unknown characters. This rock is by itself and not a portion of the Mountains; it is of White Flint; ranges North and South; is from twelve to fourteen feet long; and from eight to ten wide at base and top; and of an uneven surface. On the top I did not perceive any characters; but the sides all around are irregularly charged with unknown characters, made not indeed with the incision of a chisel, yet most certainly with an iron tool, and that by pecks or picking, after the manner of the Dighton Rock. The Lacunae or excavations are from a quarter to an inch wide; and from one tenth to two tenths of an inch deep. The engraving did not appear to be recent or new, but very old.” GEORGIA.
- 64. Charles C. Jones, jr., (a) describes a petroglyph in Georgia as follows: In Forsyth county, Georgia, is a carved or incised bowlder of fine grained granite, about 9 feet long, 4 feet 6 inches high, and 3 feet broad at its widest point. The figures are cut in the bowlder from one-half to three-fourths of an inch deep. It is generally believed that they are the work of the Cherokees. The illustration given by him is here reproduced in Fig. 38. It will be noted that the characters in it are chiefly circles, including plain, nucleated, and concentric, sometimes two or more being joined by straight lines, forming what is now known as the “spectacle shaped” figure. The illustrations should be compared with the many others presented in this paper under the heading of Cup Sculptures, see Chapter v, infra. Fig. 38.—Petroglyphs in Georgia.
- 65. Dr. M. F. Stephenson (a) mentions sculptures of human feet, various animals, bear tracks, etc., in Enchanted mountain, Union county, Georgia. The whole number of sculptures is reported as one hundred and forty-six. Mr. Jones (b) gives a different résumé of the objects depicted, as follows:
- 66. Upon the Enchanted mountain, in Union county, cut in plutonic rock, are the tracks of men, women, children, deer, bears, bisons, turkeys, and terrapins, and the outlines of a snake, of two deer, and of a human hand. These sculptures—so far as they have been ascertained and counted—number one hundred and thirty-six. The most extravagant among them is that known as the footprint of the “Great Warrior.” It measures 18 inches in length and has six toes. The other human tracks and those of the animals are delineated with commendable fidelity. IDAHO. Mr. G. K. Gilbert, of the U. S. Geological Survey, has furnished a small collection of drawings of Shoshonean petroglyphs from Oneida, Idaho, shown in Fig. 39. Some of them appear to be totemic characters, and possibly were made to record the names of visitors to the locality.
- 67. Fig. 39.—Petroglyphs in Idaho (Shoshonean). Mr. Willard D. Johnson, of the U. S. Geological Survey, reports pictographic remains observed by him near Oneida, Idaho, in 1879. The figures represent human beings and were on a rock of basalt. A copy of another petroglyph found in Idaho appears in Fig. 1092, infra. ILLINOIS. Petroglyphs are reported by Mr. John Criley as occurring near Ava, Jackson county, Illinois. The outlines of the characters observed by him were drawn from memory and submitted to Mr. Charles S. Mason, of Toledo, Ohio, through whom they were furnished to the
- 68. Bureau of Ethnology. Little reliance can be placed upon the accuracy of such drawing, but from the general appearance of the sketches the originals of which they are copies were probably made by one of the middle Algonquian tribes of Indians. The “Piasa” rock, as it is generally designated, was referred to by the missionary explorer Marquette in 1675. Its situation was immediately above the city of Alton, Illinois. Marquette’s remarks are translated by Dr. Francis Parkman (a) as follows: On the flat face of a high rock were painted, in red, black, and green, a pair of monsters, each “as large as a calf, with horns like a deer, red eyes, a beard like a tiger, and a frightful expression of countenance. The face is something like that of a man, the body covered with scales; and the tail so long that it passes entirely round the body, over the head, and between the legs, ending like that of a fish.” Another version, by Davidson and Struvé (a), of the discovery of the petroglyph is as follows: Again they (Joliet and Marquette) were floating on the broad bosom of the unknown stream. Passing the mouth of the Illinois, they soon fell into the shadow of a tall promontory, and with great astonishment beheld the representation of two monsters painted on its lofty limestone front. According to Marquette, each of these frightful figures had the face of a man, the horns of a deer, the beard of a tiger, and the tail of a fish so long that it passed around the body, over the head, and between the legs. It was an object of Indian worship and greatly impressed the mind of the pious missionary with the necessity of substituting for this monstrous idolatry the worship of the true God. A footnote connected with the foregoing quotation gives the following description of the same rock:
- 69. Near the mouth of the Piasa creek, on the bluff, there is a smooth rock in a cavernous cleft, under an overhanging cliff, on whose face, 50 feet from the base, are painted some ancient pictures or hieroglyphics, of great interest to the curious. They are placed in a horizontal line from east to west, representing men, plants, and animals. The paintings, though protected from dampness and storms, are in great part destroyed, marred by portions of the rock becoming detached and falling down. Mr. McAdams (a), of Alton, Illinois, says “The name Piasa is Indian and signifies, in the Illini, ‘The bird which devours men.’” He furnishes a spirited pen-and-ink sketch, 12 by 15 inches in size and purporting to represent the ancient painting described by Marquette. On the picture is inscribed the following in ink: “Made by Wm. Dennis, April 3d, 1825.” The date is in both letters and figures. On the top of the picture in large letters are the two words, “FLYING DRAGON.” This picture, which has been kept in the old Gilham family of Madison county and bears the evidence of its age, is reproduced as Fig. 40. Fig. 40.—The Piasa petroglyph. He also publishes another representation (Fig. 41) with the following remarks: One of the most satisfactory pictures of the Piasa we have ever seen is in an old German publication entitled “The
- 70. Valley of the Mississippi Illustrated. Eighty illustrations from nature, by H. Lewis, from the Falls of St. Anthony to the Gulf of Mexico,” published about the year 1839 by Arenz Co., Düsseldorf, Germany. One of the large full- page plates in this work gives a fine view of the bluff at Alton, with the figure of the Piasa on the face of the rock. It is represented to have been taken on the spot by artists from Germany. We reproduce that part of the bluff (the whole picture being too large for this work) which shows the pictographs. In the German picture there is shown just behind the rather dim outlines of the second face a ragged crevice, as though of a fracture. Part of the bluff’s face might have fallen and thus nearly destroyed one of the monsters, for in later years writers speak of but one figure. The whole face of the bluff was quarried away in 1846-’47. Fig. 41.—The Piasa petroglyph. Under Myths and Mythic Animals, Chapter xiv, Section 2, are illustrations and descriptions which should be compared with these accounts, and Chapter xxii gives other examples of errors and discrepancies in the description and copying of petroglyphs. Mr. A. D. Jones (a) says of the same petroglyph:
- 71. After the distribution of firearms among the Indians, bullets were substituted for arrows, and even to this day no savage presumes to pass the spot without discharging his rifle and raising his shout of triumph. I visited the spot in June (1838) and examined the image and the ten thousand bullet marks on the cliff seemed to corroborate the tradition related to me in the neighborhood. Mr. McAdams, loc. cit., also reports regarding Fig. 42: Fig. 42.—Petroglyph on the Illinois river. Some twenty-five or thirty miles above the mouth of the Illinois river, on the west bank of that stream, high up on the smooth face of an overhanging cliff, is another interesting pictograph sculptured deeply in the hard rock. It remains to-day probably in nearly the same condition it was when the French voyagers first descended the river and got their first view of the Mississippi. The animal-like body, with the human head, is carved in the rock in outline. The huge eyes are depressions like saucers, an inch or more in depth, and the outline of the body has been scooped out in the same way; also the mouth. The figure of the archer with the drawn bow, however, is painted, or rather stained with a reddish brown pigment, over the sculptured outline of the monster’s face.
- 72. Mr. McAdams suggests that the painted figure of the human form with the bow and arrows was made later than the sculpture. The same author (b) says, describing Fig. 43: Fig. 43.—Petroglyph near Alton, Illinois. Some 3 or 4 miles above Alton, high up beneath the overhanging cliff, which forms a sort of cave shelter on the smooth face of a thick ledge of rock, is a series of paintings, twelve in number. They are painted or rather stained in the rock with a reddish brown pigment that seems to defy the tooth of time. It may be said, however, that their position is so sheltered that they remain almost perfectly dry. We made sketches of them some thirty years ago and on a recent visit could see that they had changed but little, although their appearance denotes great age. These pictographs are situated on the cliff more than a hundred feet above the river. A protruding ledge, which is easily reached from a hollow in the bluff, leads to the cavernous place in the rock. Mr. James D. Middleton, formerly of the Bureau of Ethnology, mentions the occurrence of petroglyphs on the bluffs of the Mississippi river, in Jackson county, about 12 miles below Rockwood. Also of others about 4 or 5 miles from Prairie du Rocher, near the Mississippi river. IOWA. Mr. P. W. Norris, of the Bureau of Ethnology, found numerous caves on the banks of the Mississippi river, in northeastern Iowa, 4 miles
- 73. south of New Albion, containing incised petroglyphs. Fifteen miles south of this locality paintings occur on the cliffs. He also discovered painted characters upon the cliffs on the Mississippi river, 19 miles below New Albion. KANSAS. Mr. Edward Miller reports in Proceedings of the American Philosophical Society, vol. x, 1869, p. 383, the discovery of a petroglyph near the line of the Union Pacific railroad, 15 miles southeast of Fort Harker, formerly known as Fort Ellsworth, Kansas. The petroglyph is upon a formation belonging to No. 1, Lower Cretaceous group, according to the classification of Meek and Hayden. The parts of the two plates vii and viii of the work cited, which bear the inscriptions, are now presented as Fig. 44, being from two views of the same rock. Fig. 44.—Petroglyphs in Kansas. KENTUCKY. Mr. James D. Middleton, formerly of the Bureau of Ethnology, in a letter dated August 14, 1886, reports that at a point in Union county, Kentucky, nearly opposite Shawneetown, Illinois, petroglyphs are
- 74. found, and from the description given by him they appear to resemble those in Jackson county, Illinois, mentioned above. Mr. W. E. Barton, of Wellington, Ohio, in a communication dated October 4, 1890, writes as follows: At Clover Bottom, Kentucky, on a spur of the Big Hill, in Jackson county, about 13 miles from Berea, is a large rock which old settlers say was covered with soil and vegetation within their memory. Upon it are representations of human tracks, with what appear to be those of a bear, a horse, and a dog. These are all in the same direction, as though a man leading a horse, followed the dog upon the bear’s track. Crossing these is a series of tracks of another and larger sort which I can not attempt to identify. The stone is a sandstone in the subcarboniferous. As I remember, the strata are nearly horizontal, but erosion has made the surface a slope of about 20°. The tracks ascending the slope cross the strata. I have not seen them for some years. The crossing of the strata shows that the tracks are the work of human hands, if indeed it were not preposterous to think of anything else in rocks of that period. Still the tracks are so well made that one is tempted to ask if they can be real. They alternate right and left, though the erosion and travel have worn out some of the left tracks. A wagon road passes over the rock and was the cause of the present exposure of the stone. It can be readily found a fourth of a mile or less from the Pine Grove schoolhouse. MAINE. A number of inscribed rocks have been found in Maine and information of others has been obtained. The most interesting of
- 75. them and the largest group series yet discovered in New England is shown in Pl. xii. BUREAU OF ETHNOLOGY TENTH ANNUAL REPORT PL. XII PETROGLYPHS IN MAINE. The rock upon which the glyphs appear is in the town of Machiasport, Maine, at Clarks point, on the northwestern side of Machias bay, 2 miles below the mouth of Machias river. The rock or ledge is about 50 feet long from east to west and about fifteen feet in width, nearly horizontal for two-thirds its length, from the bank or western end at high water, thence inclining at an angle of 15° to low-water mark. Its southern face is inclined about 40°. The formation is schistose slate, having a transverse vein of trap dike extending nearly across its section. Nearly the entire ledge is of blue-black color, very dense and hard except at the upper or western end, where the periodical formation of ice has scaled off thin layers of surface and destroyed many figures which are remembered by persons now living. The ebb and flow of tides, the abrasion of moving beach stones or pebble wash and of ice-worn bowlders, have also effaced many figures along the southern side, until now but one or two indentations are discernible. Visitors, in seeking to remove some portion of the rock as a curiosity or in striving to perpetuate their initials, have obscured several of the most interesting, and until recently the best defined figures. It was also evident to the present writer, who carefully examined the rock in 1888, that it lay much deeper in the water than once had been the
- 76. case. At the lowest tides there were markings seen still lower, which could not readily have been made if that part of the surface had not been continuously exposed. The depression of a rock of such great size, which was so gradual that it had not been observed by the inhabitants of the neighboring settlement, is an evidence of the antiquity of the peckings. The intaglio carving of all the figures was apparently made by repeated blows of a pointed instrument—doubtless of hard stone; not held as a chisel, but working by a repetition of hammerings or peckings. The deepest now seen is about three-eighths of an inch. The amount of patient labor bestowed upon these figures must have been great, considering the hardness of the rock and the rude implement with which they were wrought. There is no extrinsic evidence of their age. The place was known to traders early in the seventeenth century, and much earlier was visited by Basque fishermen, and perhaps by the unfortunate Cortereals in 1500 and 1503. The descendants of the Mechises Indians, a tribal branch of the Abnaki, who once occupied the territory between the St. Croix and Narraguagus rivers, when questioned many years ago, would reply in substance that “all their old men knew of them,” either by having seen them or by traditions handed down through many generations. Several years ago Mr. H. R. Taylor, of Machias, who made the original sketch in 1868 and kindly furnished it to the Bureau of Ethnology, applied to a resident Indian there (Peter Benoit, then nearly 80 years old) for assistance in deciphering the characters. He gave little information, but pointed out that the figures must not all be read “from one side only,” thus, the one near the center of the sketch, which seen from the south was without significance, became from the opposite point a squaw with sea fowl on her head, denoting, as he said, “that squaw had smashed canoe, saved beaver-skin, walked one-half moon all alone toward east, just same as heron wading alongshore.” Also that the three lines below the figure mentioned, which together resemble a bird track or a trident, represent the
- 77. three rivers, the East, West, and Middle rivers of Machias, which join not far above the locality. The mark having a rough resemblance to a feather, next on the right of this river-sign, is a fissure in the rock. Most of the figures of human beings and other animals are easily recognizable. Peckings of a character similar to those on the Picture rock at Clarks point, above described, were found and copied 600 feet south of it at high-water mark on a rock near Birch point. Others were discovered and traced on a rock on Hog island, in Holmes bay, a part of Machias bay. All these petroglyphs were without doubt of Abnaki origin, either of the Penobscot or the Passamaquoddy divisions of that body of Indians. The rocks lay on the common line of water communication between those divisions and were convenient as halting places. MARYLAND. In the Susquehanna river, about half a mile south of the state line, is a group of rocks, several of the most conspicuous being designated as the “Bald Friars.” Near by are several mound-shaped bowlders of the so-called “nigger-head” rock, which is reported as a dark- greenish chlorite schist. Upon the several bowlders are deep sculpturings, apparently finished by rubbing the depression with stone, or wood and sand, thus leaving sharp and distinct edges to the outlines. Some of these figures are an inch in depth, though the greater number are becoming more and more eroded by the frequent freshets, and by the running ice during the breaking up in early spring of the frozen river. The following account is given by Prof. P. Frazer (a): Passing the Pennsylvania state line one reaches the southern barren serpentine rocks, which are in general tolerably level for a considerable distance.
- 78. About 700 yards, or 640 meters, south of the line, on the river shore, are rocks which have been named the Bald Friars. French’s tavern is here, at the mouth of a small stream which empties into the Susquehanna. About 874 yards (800 meters) south of this tavern are a number of islands which have local names, but which are curious as containing inscriptions of the aborigines. The material of which most of these islands are composed is chlorite schist, but as this rock is almost always distinguished by the quartz veins which intersect it, so in this case some of the islands are composed of this material almost exclusively, which gives them a very striking white appearance. One of these, containing the principal inscriptions, is called Miles island. The figures, which covered every part of the rocks that were exposed, were apparently of historical or at least narrative purport, since they seemed to be connected. Doubtless the larger portion of the inscription has been carried away by the successive vicissitudes which have broken up and defaced, and in some instances obliterated, parts of which we find evidence of the previous existence on the islands. Every large bowlder seems to contain some traces of previous inscription, and in many instances the pictured side of the bowlder is on its under side, showing that it has been detached from its original place. The natural agencies are quite sufficient to account for any amount of this kind of displacement, for the rocks in their present condition are not refractory and offer no great resistance to the wear of weather and ice; but in addition to this must be added human agencies.
- 79. Welcome to our website – the perfect destination for book lovers and knowledge seekers. We believe that every book holds a new world, offering opportunities for learning, discovery, and personal growth. That’s why we are dedicated to bringing you a diverse collection of books, ranging from classic literature and specialized publications to self-development guides and children's books. More than just a book-buying platform, we strive to be a bridge connecting you with timeless cultural and intellectual values. With an elegant, user-friendly interface and a smart search system, you can quickly find the books that best suit your interests. Additionally, our special promotions and home delivery services help you save time and fully enjoy the joy of reading. Join us on a journey of knowledge exploration, passion nurturing, and personal growth every day! ebookbell.com