Prognostics with optimization

Reducing Hospital Readmissions by Integrating
Empirical Prediction with Resource Optimization
Jonathan E. Helm
Operations and Decision Technologies, Kelley School of Business, Indiana University, 1309 E. Tenth Street, Bloomington, Indiana 47405,
USA, helmj@indiana.edu
Adel Alaeddini
Department of Mechanical Engineering, University of Texas, San Antonio One UTSA Circle, San Antonio, Texas 78249, USA
adel.alaeddini@utsa.edu
Jon M. Stauffer, Kurt M. Bretthauer
Operations and Decision Technologies, Kelley School of Business, Indiana University, 1309 E. Tenth Street, Bloomington, Indiana 47405,
USA, stauffer@indiana.edu, kbrettha@indiana.edu
Ted A. Skolarus
Department of Urology, University of Michigan, VA, Health Services Research & Development (HSR&D) Center for Clinical Management
Research, VA, Ann Arbor Healthcare System, 3875 Taubman Center, 1500 East Medical Center Drive, Ann Arbor, Michigan 48109, USA
tskolar@med.umich.edu
Hospital readmissions present an increasingly important challenge for health-care organizations. Readmissions are
expensive and often unnecessary, putting patients at risk and costing $15 billion annually in the United States
alone. Currently, 17% of Medicare patients are readmitted to a hospital within 30 days of initial discharge with readmis-
sions typically being more expensive than the original visit to the hospital. Recent legislation penalizes organizations with
a high readmission rate. The medical literature conjectures that many readmissions can be avoided or mitigated by post-
discharge monitoring. To develop a good monitoring plan it is critical to anticipate the timing of a potential readmission
and to effectively monitor the patient for readmission causing conditions based on that knowledge. This research develops
new methods to empirically generate an individualized estimate of the time to readmission density function and then uses
this density to optimize a post-discharge monitoring schedule and staffing plan to support monitoring needs. Our
approach integrates classical prediction models with machine learning and transfer learning to develop an empirical den-
sity that is personalized to each patient. We then transform an intractable monitoring plan optimization with stochastic
discharges and health state evolution based on delay-time models into a weakly coupled network flow model with tracta-
ble subproblems after applying a new pruning method that leverages the problem structure. Using this multi-methodolog-
ic approach on two large inpatient datasets, we show that optimal readmission prediction and monitoring plans can
identify and mitigate 40–70% of readmissions before they generate an emergency readmission.
Key words: hospital readmissions; post-discharge patient monitoring; readmission risk profiling; Bayesian survival analy-
sis; delay-time models of readmissions
History: Received: December 2013; Accepted: December 2014 by Tsan-Ming Choi, after 1 revision.
1. Introduction
Hospital and medical center readmissions is a serious
health-care issue demanding increased attention as
costs continue to rise and patient care suffers. Based
on a report to Congress in 2008, over 17% of Medicare
patients were readmitted in the first 30 days after dis-
charge, accounting for more than $15 billion dollars
per year (Foster and Harkness 2010). Not only are
readmissions expensive, recent studies have also
linked the rate of readmission to quality of care in
medical centers (e.g., Halfon et al. 2006). Surprisingly,
Foster and Harkness (2010) found that a significant
percent of readmissions are avoidable through better
post-discharge management; of the $15 billion spent,
$12 billion was associated with potentially preventable
readmissions. Current strategies to reduce readmis-
sions focus on (i) identifying high-risk patients (e.g.,
Kansagara et al. 2011, Rosenberg et al. 2007, Wall-
mann et al. 2013), or (ii) developing an effective plan
for post-discharge care (e.g., Jack et al. 2009). While
these heuristic clinical approaches have proven effec-
tive in avoiding readmissions, there remains signifi-
cant opportunity for an approach that combines
233
Vol. 25, No. 2, February 2016, pp. 233–257 DOI 10.1111/poms.12377
ISSN 1059-1478|EISSN 1937-5956|16|2502|0233 © 2015 Production and Operations Management Society

rigorous empirical modeling to predict time to read-
mission with optimization to design schedules and
allocate staff for post-discharge monitoring. To have
the largest possible impact on readmissions, it is nec-
essary to know both when a patient is likely to be re-
admitted (empirical prediction model) and when to
monitor that patient to identify the condition before it
triggers a readmission (optimization model). This
study represents a multi-methodology effort aimed at
integrating clinical, statistical, and operations man-
agement techniques to (i) quantify post-discharge risk
of readmission for each patient over time, (ii) to
design optimal post-discharge treatment plans for
early detection and avoidance of potential readmis-
sions, and (iii) to allocate sufficient system capacity to
be able to administer the optimal treatment plans for
a cohort of patients.
Numerous efforts have focused on capturing the
key dynamics of the readmission system (Desai et al.
2009, Kansagara et al. 2011). The study of readmission
risk factors typically falls into three major categories:
(i) patient attributes such as history of readmission,
severity of illness, comorbidity, age, gender, life satis-
faction, change in clinical variables, source of pay-
ment, etc. (e.g., Dunlay et al. 2009, Wallmann et al.
2013, Watson et al. 2011); (ii) factors targeting the pre-
discharge process including length of stay, adequacy
of discharge plan, nursing environment of the hospi-
tal, characteristics of the physician, etc., (e.g., McHugh
and Ma 2013, Rosen et al. 2013); and finally (iii) fac-
tors targeting the post-discharge process including
inadequacy of post-discharge planning and follow
up, non-compliance with medication and diet, failed
social support, impairment of self-care, etc. (e.g., Her-
nandez et al. 2010, Wallmann et al. 2013, Watson
et al. 2011). Using the above risk factors a number of
health-care systems have started implementing online
readmission risk calculators. Some of these calculators
may be found at http://riskcalc.sts.org/STSWebRisk-
Calc273/, by the Society of Thoracic Surgeons which
predicts the risk of operative mortality and morbidity
after adult cardiac surgery, and at http://www.read-
missionscore.org, by the Center for Outcomes
Research and Evaluation (CORE), which helps predict
a patient’s likelihood of readmission for heart failure
within 30 days of discharge. Despite their benefits,
these calculators have serious limitations. They (i)
assume homogeneity of the population and hospital’s
performance; (ii) provide no estimate on time to read-
mission; and (iii) provide no guidance on how to use
the estimates to make better care decisions. Our meth-
ods will address these deficiencies.
Recently, researchers have begun to investigate the
impact of targeted discharge planning and post-
discharge management on reducing readmissions,
focusing on financial incentives/cost-effectiveness,
pre-discharge patient education, and improved post-
discharge management. In particular, several studies
claim that post-discharge management can reduce
readmissions by 12% to 30% (see Gonseth et al. 2004)
and as high as 85% (see Fonarow et al. 1997) by target-
ing high-risk populations (see Minott 2008, Wolinsky
et al. 2009), telemonitoring (see Graham et al. 2012),
and other monitoring strategies. By integrating
patient risk calculations and empirical predictions of
time to readmission with optimization methods to
design monitoring plans, we capture both of the
high-impact approaches (risk profiling and planned
monitoring) from the medical literature in a quantitative
framework for optimally designing these post-discharge
monitoring plans that are currently designed using
expert judgment or ad hoc approaches.
Figure 1 provides a high-level overview of our
multi-methodology approach which uses both read-
mission prediction and follow-up schedule optimi-
zation to reduce readmissions. First the Empirical
Prediction Model utilizes individual patient data to
determine a probability of readmission and
expected time to readmission for each patient. These
patient and procedure specific readmission curves
are then aggregated with K-mean clustering (or
other methods) into several different risk profiles.
The aggregated readmission curves for each risk
profile, organization-specific resource capacity and
Individual Patient Data:
• Demographic Data
• Health History
• Hospital Stay Data
Individual Patient Estimates:
• Probability of Readmission
• Time to Readmission
Patient Readmission
Costs and Savings
Patient Follow-up Schedule
Required Hospital Resources
Empirical Prediction Model
Aggregation into Risk Profiles
Optimization Model
Hospital Follow-up
Capacity and Costs
Figure 1 Multi-Methodology Model Overview
Helm, Alaeddini, Stauffer, Bretthauer, and Skolarus: Multi-Method Readmission Reduction
234 Production and Operations Management 25(2), pp. 233–257, © 2015 Production and Operations Management Society

cost information, and medical procedure-specific re-
admission cost and savings information are all used
as inputs into the Optimization Model. The Optimi-
zation Model uses these inputs to determine an
optimal follow-up schedule for each patient risk
profile and determine the number of resources the
hospital or health system will need to execute all
expected follow-up schedules.
While previous literature relies on a siloed
approach, focusing either on predicting readmissions
or on strategies to reduce readmissions, this research
integrates the two using advanced mathematical, sta-
tistical, and operations management techniques com-
bined with clinical expertise. We not only develop an
integrative framework for investigating both aspects
of readmission modeling simultaneously, we also
contribute new methods to each of the areas. To the
best of our knowledge, existing studies have not
effectively considered heterogeneity among patient
populations, and are not able to adapt population-
based readmission estimates to individual patients.
Further, previous readmission prediction models
have only focused on small groups of patients with a
single readmission triggering condition (e.g., elderly
cardiovascular patients), and the results are often not
generalizable to other cases (see Feudtner et al. 2009,
Gonseth et al. 2004). In addition, most of the available
studies have not effectively used the array of avail-
able machine-learning techniques to improve their
results. This study addresses these deficiencies by
enabling individualized readmission probability esti-
mates and a generalizable method that can encom-
pass diverse patient populations and multiple
readmission causing conditions over an arbitrary time
period. Further, existing models have been lacking
comprehensive optimization approaches to design
tailored post-discharge management plans. No litera-
ture to our knowledge captures, as we intend to do,
the health-care organization’s ability to support a
large-scale implementation of a post-discharge
management scheme that simultaneously solves for
post-discharge monitoring timing and the organiza-
tional resource capacity needed to implement such
schedules.
Finally, we demonstrate how this multi-methodol-
ogy approach can be applied via an extensive case
study and numerical analysis using two different
datasets from (i) a partner hospital in Michigan
including 2449 patients with 17 diagnoses, 3108 read-
missions, and 15 demographic, socioeconomic, and
clinical factors, etc. (ii) the State Inpatient Databases
(SID) for 5000 patients diagnosed with bladder, kid-
ney, and prostate cancer in 2009 along with other can-
cers (see http://www.hcup-us.ahrq.gov/db/state/
siddist/SID_Introduction.jsp). The results for the two
datasets were structurally similar, so for the purposes
of cohesive exposition we focus on the results from
the partner hospital in Michigan for this study.
Section 2 develops the empirical model to predict
readmission occurrence and timing. Section 3 uses the
predicted empirical readmission density from section
2 to develop a follow-up schedule for patients and
staffing plan for a follow-up organization. Section 4
brings both components together, empirical predic-
tion and resource optimization, in a case study using
historical inpatient readmission data to design a prac-
tical post-discharge monitoring schedule and gener-
ate insights into tactical and operational management
of post-discharge care. These results confirm the con-
jecture in the medical literature that between 12% and
85% of readmissions can be avoided or identified
early through better post-discharge plans and show
how to effectively design such plans. Section 5 con-
cludes the study.
2. Stage 1: Empirical Modeling to
Predict Time to Readmission
While a number of studies have focused on predicting
whether or not a patient will be readmitted within
30 days (see van Walraven et al. 2010), there is only
one article to our knowledge that focuses on predict-
ing the time to readmission (see Yu et al. 2013). While
Yu et al. (2013) shares similarities with our work,
there are important differences in the two approaches.
From a methodological perspective Yu et al. (2013),
among other studies, does not consider which condi-
tion has caused the readmission, for example, infec-
tion, dehydration, kidney failure etc. We are able to
capture this feature using a frailty approach to model
these conditions as latent competing risks with sto-
chastic dependence. In addition, Yu et al. (2013),
among others, use a population-based approach
based on equally weighted readmission records from
a specific hospital to calculate the risk of readmission
for that hospital’s patients. However, we employ
transfer learning to weight the readmission records in
the dataset based on their similarity to the readmis-
sion record(s) for the patient of interest to: (i) further
personalize the estimate and (ii) alleviate the problem
of data scarcity. Finally, Yu et al. (2013), along with
other readmission prediction models, gives the same
importance to all readmission records regardless of
how recently the readmission occurred. Our method
assigns importance (weight) to the admission/read-
mission records based on record recency (more recent
records get more weight) using an optimization
process to choose the appropriate weights. This
accounts for the phenomenon that each patient’s
health status and/or behaviors can change over time.
From the specific modeling perspective, we use a
Bayesian approach while Yu et al. (2013) uses a
Production and Operations Management 25(2), pp. 233–257, © 2015 Production and Operations Management Society 235

classical approach. Further, we employ a parsimoni-
ous prior while Yu et al. (2013) employs a forward
selection procedure for identifying the most important
variables.
Understanding the time to readmission is critical to
making clinically effective decisions to mitigate
potential readmissions, such as when to follow up
with a patient who has been discharged from the hos-
pital. In this section, we develop empirical prediction
models to accurately capture the probability distribu-
tion on time to readmission based on two different
datasets (the State Inpatient Database (SID) as well as
a dataset from a partner hospital in Michigan) to
show that our methods can be used broadly (e.g., on
SID) or tailored to a specific hospital.
Beyond exploring the new area of predicting time
to readmission, we also address two other features
that are prevalent in health care: (i) the need to per-
sonalize the prediction method to each individual
patient, and (ii) scarcity of relevant data. The result of
the empirical modeling in this section is a set of tai-
lored probability distributions (one for each individ-
ual patient) that is personalized for each patient in
our datasets. Our approach builds up the prediction
model through three steps as shown in Figure 2. Step
1 (section 2.1) develops a general population estimate
for time to readmission, which accounts for demo-
graphic, socioeconomic, health history, co-morbidity,
the hospital the patient was treated at, and other rele-
vant patient and system characteristics. In section 2.1,
we also discuss how we are able to incorporate the
cause of readmission into our prediction model. We
do so by developing a Weibull regression model that
incorporates observable and unobservable risk factors.
The model from Step 1 is then personalized in Step
2 (section 2.2) by parameterizing the Weibull model
using each patient’s personal history of hospital
admissions and readmissions to date. Step 3 (also sec-
tion 2.2) addresses the problem of data scarcity, which
occurs when an individual has too few prior records
to adequately parameterize the model with their indi-
vidual data alone (Step 2) and when applying the
method to a new hospital or group that has little rele-
vant data. For example, when personalizing the read-
mission estimate, we are able to use data from all
patients in our dataset (not just from the patient
whose time to readmission curve is currently being
estimated), by adding weights to the data records.
Higher weights indicate a higher level of statistical
similarity of any given patient in the dataset with the
target patient. This approach, called transfer learning,
and the specifics of calculating and incorporating
weights are discussed in section 2.2.
In section 2.3 we discuss the methods and algo-
rithms used to apply the approaches in sections 2.1
and 2.2 to our real-world datasets. While the data
about a particular individual or specific hospital may
be small, the overall dataset we intend the model to
work with will be large. With large datasets, the com-
mon methods for prediction have significant draw-
backs. For example, machine learning often suffers
from results being difficult to interpret and sometimes
yields patterns that are a product of random fluctua-
tions, while more classical prediction models employ
oversimplifying assumptions that lead to incorrect
conclusions. To overcome these limitations, we
develop an empirical prediction model that integrates
both classical prediction methods with machine learn-
ing using a Bayesian framework. We conclude the
section by comparing the accuracy of our prediction
model against other commonly used prediction mod-
els in the literature.
2.1. Population-Based Model of Time to
Readmission
We begin by building a population-based estimate of
time to readmission in which we consider the impact
of (i) time after initial discharge from the hospital, (ii)
patient-specific risk factors impacting likelihood of re-
admission, and (iii) unobservable or random effects
that capture patient heterogeneity. We capture time to
readmission using a Weibull regression model. We
begin with a hazard rate function, h(t). In our optimi-
zation model presented in section 3, this hazard rate
can be used to model the deterministic arrival rate
function of a non-homogeneous Poisson process
(NHPP) capturing the arrival of readmission-causing
failures (as is common in delay-time analysis), how-
ever, the NHPP assumption is not necessary for esti-
mation of the survival model. The probability that a
patient has not yet been readmitted by time t is there-
fore given by SðtÞ ¼ exp À
Rt
0 hðuÞdu

, which we call
the survival function. The hazard function, however,
Data
Source
All Patients: Demographic, Socio-Economic,
Health History, Comorbidity, etc.
Individual patient’s re/admission records
Purpose
Generalized correlated frailty model
Population Estimate of Readmissions
Personalized Prediction for each
Individual Patient
Increasing Prediction Accuracy by
Including Data from Similar Patients
Similarity among patients’ readmission records
Bayesian Inference, Markov Chain Monte Carlo
Methods
Local regression similarity index, Bayesian
weighting, titled time framing
Step 1 Step 3Step 2
Methods
Figure 2 Framework for Predicting Patient Readmissions

depends not only on time but also on a set of K risk
factors for readmission, X ¼ ½x1; . . .; xKŠ. We consider
the following factors available to us in the data: length
of stay, gender, age, employment status, insurance
coverage level, profession/military rank, ward(s) vis-
ited during inpatient stay, principal diagnosis, and
source of admission, that is, VA hospital, nursing
home, home, non-VA hospital. To tailor our hazard
rate function to these patient characteristics, we
employ a Weibull regression model which incorpo-
rates the important risk factors that affect probability
of readmission as follows:
hðtjXÞ ¼ h0ðtÞ Á expðXBÞ ; ð1Þ
where h0ðtÞ ¼ qtqÀ1
is a Weibull function.
B ¼ ½b0; b1; . . .; bKŠ0
is a vector of K regression
parameters (risk factor coefficients) to be estimated.
However, not all of the risk factors affecting Equa-
tion 1 are easily known or even measurable. For
example, patients can be readmitted for several dif-
ferent post-discharge complications—common ones
include infection, dehydration, kidney failure, fail-
ure to thrive—where these conditions are all
competing to cause a readmission, may exhibit sto-
chastic dependence, and are not observable at the
time a prediction is made. In the data, we are only
able to observe the factor that caused the readmis-
sion, for example, infection, which is essentially the
minimum failure time of all the latent risk factors
that could cause readmission. To account for such
latent competing risks and their stochastic depen-
dence, we use a “frailty” approach to extend
the Weibull regression model (see Clayton 1978,
Hougaard and Hougaard 2000, Oakes 1989). If
there are M frailty terms, m1; . . .; mM, corresponding
to the M latent risks, then the risk-specific hazard
rate for the mth latent risk factor can be written as
follows:
hmðtjX; mmÞ ¼ h0;mðtÞ Á expðXBm þ mmÞ : ð2Þ
Equation 2 is a generalization of Equation 1 to
incorporate unmeasurable risk factors, m. Thus, we
have a different hazard rate function for each of the
competing risks that might cause a readmission—
for example, h1ðtjX; m1Þ could be the hazard rate for
infection, h2ðtjX; m2Þ could be the hazard rate for fail-
ure to thrive, etc. This allows the model to capture
not only the time to readmission but also different
time to readmission dynamics for different causes
of readmission. This could potentially help clini-
cians better target diagnostic questioning and tests
to look for specific readmission causing conditions
at different times after discharge; an idea which is
supported by the clinical literature (see Hu et al.
2014). Assuming that the vector of frailties m is
drawn from a multivariate distribution with density
gðm1; . . .; mMÞ and ti for i = 1, . . ., M is the failure
time for the ith readmission causing condition,
then the unconditional (expected) survivor function
can be calculated by integrating with respect to
density g:
Sðt1;...;tMjXÞ¼
Z
...
Z
Sðt1;...;tMjm1;...;mMÞ
Âgðm1;...;mMÞdm1...dmM
¼
Z
...
Z
PM
m¼1
È
exp½À
Z tm
0
hmðujX;mmÞduŠ
É
gðm1;...;mMÞdm1...dmM
¼
Z
...
Z
PM
m¼1
È
exp½Àtqm
m expðXBm þmmÞŠ
É
gðm1;...;mMÞdm1...dmM:
ð3Þ
The first line takes the expectation of the joint survi-
vor function over the frailty terms m1; . . .; mm. The
second line follows from the assumption made in
the frailty literature that, conditional on the frailty,
the risks for the different causes of readmis-
sion are independent (see Gordon 2002). Thus, the
joint distribution of the times to readmission from
each cause, t1; . . .; tm, decomposes into the prod-
uct of the marginal survival functions, SmðtmÞ ¼
exp½À
Rtm
0 hmðujX; mmÞduŠ. The third line follows by
integrating hmðtjX; mmÞ from Equation 2, and the fact
that, for our Weibull formulation
Rt
0 h0;mðuÞdu ¼ t
qm
m .
Recalling that in a survival model, the density func-
tion is given by f(t) = h(t)S(t), from Equations 2
and 3 the unconditional density function can be
calculated as:
fðt1; . . .; tMjXÞ ¼
Z
. . .
Z
fðt1; . . .; tMjm1; . . .; mMÞ
Â gðm1; . . .; mMÞdm1. . .dmM:
¼
Z
. . .
Z
hðt1; . . .; tMjm1; . . .; mMÞ
Â Sðt1; . . .; tMjm1; . . .; mMÞ
¼
Z
. . .
Z
PM
m¼1
È
qmtqmÀ1
m Á
Â expðXBm þ mmÞ exp½ÀexpðXBm þ mmÞtqm
m Š
É
ð4Þ
The first line follows by taking the expectation over
the frailties as in Equation 3. The second line follows
by applying the definition of f, that is, f(t) = h(t) 9 S(t).

Thus, the final result is just a product of the hazard
functions for each latent factor, hmðtmjX; mmÞ, with the
conditional survivor function again by applying the
assumption of independence of risks conditional on
the frailty terms.
The final step is to calculate the marginal likelihood
function for estimating the unknown parameters of
the model. First, as is common in health-care data, we
must account for the fact that some of the data we
have obtained will be censored. In the case of read-
missions, a logical choice of censoring limit is 30 days,
given that the current policy only penalizes readmis-
sions within 30 days. To incorporate censoring, let dim
represent the censoring indicator for the individual
i = 1, . . ., n and frailty m = 1, . . ., M that is one if the
data are uncensored and zero if the data are censored.
The following presents the standard form of the likeli-
hood function with censoring, L(B,q,m), that takes the
form of the density, f, when data are uncensored, and
the form of the survival function, S, when the data are
censored:
LðB;q;mjdataÞ¼
¼Pn
i¼1
Z
...
Z
PM
m¼1ðqmt
qmÀ1
im ÁexpðXBm þmmÞ
Âexp½ÀexpðXBm þmmÞt
qm
im ŠÞdim
Áðexp½Àt
qm
im expðXBm þmmÞŠÞð1ÀdimÞ
Âgðm1;...;mMÞdm1...dmM
¼Pn
i¼1
Z
...
Z
PM
m¼1
È
ðqmt
qmÀ1
im expðXBm þmmÞÞdim
Á
ÂexpðÀt
qm
im expðXBm þmmÞ
É
gðm1;...;mMÞdm1...dmM:
ð5Þ
where tim is the time of readmission due to latent
factor m (or the censoring time, e.g., day 30) for data
record (patient) i. The population model presented
in Equations 2–5 has the nice properties of being
able to account for the impact of time, patient obser-
vable risk factors, and unobservable competing risks
(e.g., infection, failure to thrive, dehydration, etc.)
on readmissions by estimating the unknown param-
eters from data.
2.1.1. Parameter Estimation for Time to
Readmission. The next step is to estimate the
unknown parameters, specifically the parameter of
the marginal baseline hazard q, the coefficients of our
patient risk factors, B, and frailties, m. Unfortunately,
these parameters cannot be effectively estimated
directly using conventional likelihood maximization
methods because these methods cannot directly maxi-
mize the full likelihood, given data, and the small
sample properties of these estimators have yet to be
studied (Ibrahim et al. 2005). To avoid these pitfalls,
we transform the hazard model presented in Equa-
tions 2–5 into a Bayesian model and use Markov
Chain Monte Carlo (MCMC) methods to draw param-
eters from their posterior distributions (Gilks and
Wild 1992). Not only does the Bayesian framework
enable effective parameter estimation, it also facili-
tates our approach to addressing data scarcity issues
when personalizing the prediction to individual
patients/hospitals. This framework also permits right
censored data as presented in Equation 5, where cen-
soring occurs because we observe patients at a spe-
cific point in time and if they have not yet been
readmitted or have left the database we do not
observe their true time to readmission, but instead
observe a survival time that is smaller than their time
to readmission.
We used Winbugs software for Bayesian analysis
and MCMC sampling. To set up the model in
Winbugs, we used a rectangular data format with
separate columns to represent regular readmission
and censoring times. Individuals who are censored
are given a missing value in the vector of read-
mission times, while individuals who actually get
readmitted are given a zero in the censoring time
vector. The truncated Weibull model based on Win-
bugs built in censoring function is used to include
appropriate term(s) in the full conditional distribu-
tion (similar to the model explained in Equation 5).
Detailed instructions and examples on Bayesian
analysis of Weibull regression in censored survival
analysis using Winbugs can be found at Spiegelhalter
et al. (2003).
To transform the Weibull hazard model into a
Bayesian framework, we first need to define a distri-
bution that represents our current belief about the
parameters to be estimated, which is called a Bayesian
prior. Then, we define a likelihood function, L(params|
data), that calculates the probability that the chosen
parameters are a good representation of the data.
Finally, we multiply the prior by the likelihood func-
tion and then normalize to obtain a posterior distribu-
tion on our set of parameters that better fits the
observed data. We generate the posterior distribution
by sampling the prior distribution on parameters B, q,
and m in a Monte Carlo fashion and calculating the
posterior based on the data and the result of sampling
the prior using Gibbs Sampling algorithm (see Gilks
and Wild 1992).
To avoid over-fitting, it is important to choose an
appropriate prior distribution. In particular, we
want a prior distribution in which the mode of the
parameters are likely to be near zero while the vari-
ance is monotonically decreasing with positive finite
value at zero (see Gustafson 1997). This discourages
the model from selecting too many variables in
the data fitting process. From our numerical experi-
ments setting, the prior distribution of the risk
factors, B, to be independent double exponential

random variables, using multivariate normal distri-
bution to represent the joint distribution of the frail-
ties, mm, and choosing a gamma distribution for the
parameter of the marginal baseline hazard, qm, per-
formed well for our data. In the next section, we
show how we can employ transfer learning to
parameterize our Weibull regression model
described in this section with enough data and yet
still personalize the readmission prediction.
2.2. Personalizing Readmission Predictions and
Data Scarcity
A key feature of health-care modeling is the need to
personalize prediction and forecasting models to
account for individual patient characteristics because
aggregate population-based models do not often per-
form well on a patient-by-patient basis. Personalizing
time to readmission predictions also enables us to
develop risk profiles that allow us to tailor our moni-
toring decision framework in section 3 and signifi-
cantly outperform a population-based monitoring
plan. To obtain a population-based estimate, we sim-
ply use all the data available in our datasets in the
likelihood function, L (Equation 5), when estimating
the parameters. To tailor the estimate to one specific
patient, we would only use that patient’s data in cal-
culating the likelihood function. While the population
estimate is not discerning enough, a single patient
would not have nearly enough data points (historical
admission/readmission records) to adequately esti-
mate the large number of parameters in Equation 5.
To overcome this, we use an approach called transfer
learning (see Pan and Yang 2010), which includes
records from statistically similar patients in the likeli-
hood function to increase the amount of data we can
use to estimate the parameters for the particular indi-
vidual.
In transfer learning, we calculate the posterior
distribution on the parameters we wish to estimate,
B,q,m, using all the data but giving more influence
to data records based on how similar they are to
the patient we are trying to estimate the parameters
for. In estimating the time to readmission den-
sity function for patient i, we have a similarity/
relevance weight for patient j that is given by wi;j.
For the data from patient j, we then take the like-
lihood to the power wi;j. This way, data from a
higher weight for patient j, indicating they are
more similar to patient i, has more influence in the
likelihood function.
The weighting scheme to calculate similarity and
relevance of data records includes two factors: (i) Re-
admission record similarity: measuring how similar
two patients in the data are; and (ii) Recency of read-
mission records: giving a higher weight to more
recent admission/readmission records.
2.2.1. Record Similarity (W1). To calculate read-
mission record similarity, (W1), we first divide the
index of risk factors 1, . . ., K into the set of indices
that represent numeric factors, K1, and categorical fac-
tors, K2. Separating the factors into numeric and cate-
gorical groups and comparing only against other
factors in the same group mitigates a potential bias if
jK2j [ [ jK1j (where |Á| is the cardinality of the set)
that could be introduced because the numeric factors
will always be less than one, while the categorical fac-
tors will be either 0 or 1. Next, we use cosine similar-
ity measure (Pang-Ning et al. 2006) for calculating the
similarity among numeric factors, and simple match-
ing (Sokal 1958) for categorical factors. Next, we use
the weighted average of the numerical factors (with
weight jK1j) and categorical factors (with weight jK2j)
to provide the total similarity measure. We normalize
the numerical risk factors for patient i to fall within
the interval (0,1). Similarity for categorical risk factors
for patients i and j, xi;n; xj;n, for n 2 K2, is represented
by an indicator, 1fxi;n; xj;ng ¼ 1 if the categorical fac-
tors are identical and 0 otherwise. The total similarity
measure will then be given by:
W1ij
¼
x1W11ij
þ x2W12ij
x1 þ x2
ð6Þ
where x1 ¼ jK1j and x2 ¼ jK2j are the number of
factors in the numerical and categorical groups,
respectively. W11ij
¼
P
k2K1
xi;kÁxj;k
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP
k2K1
ðxi;kÞ2
q
Á
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP
k2K1
ðxj;kÞ2
q , and
W12ij
¼
P
k2K2
1fxi;k;xj;kg
jK2j . In other words, to calculate the
overall readmission record similarity, like factors are
compared with like factors and then averaged
according to how many of that type of factor
appears in the data. In this way, the influence of the
categorical variables on the continuous variables is
reduced.
2.2.2. Record Recency (W2). For readmission
record recency, we developed a tilted time framing mech-
anism that is closely related to exponentially weighted
moving average (EWMA) smoothing. A weighting factor
of W2 is defined based on the generalized logistic func-
tion W2ðtÞ ¼ ½1 þ Q Á expð À Aðt À CÞ
1
gÞÞŠÀ1
, where t is
the date the readmission record occurred, Γ is the
current date, and A,g,Q are the parameters of the
logistic functions determining the shape and scale of
the function (Richards 1959). These parameters can
be determined by minimizing the mean squared
error (MSE) of the estimated probabilities of read-
missions (in the training dataset), and the respective
empirical probabilities (in the validation dataset).
The total weight applied to records from patient
j being used to estimate time to readmission for

patient i, wi;j, is determined by multiplying the
similarity and recency measures.
This approach can also be used at the hospital level,
employing information from health systems with
available data on readmission rates to predict the rate
of readmissions in new medical centers or those with
insufficient data. We summarize the algorithm for
empirically estimating our time to readmission den-
sity function in Table 1.
2.3 Analyzing Model Performance on Real Data
To demonstrate the effectiveness of our approach, we
compare its predictive power to other effective pre-
diction methods from the literature including: (i) clas-
sification and regression trees (CART), (ii) multilayer
perceptron (MLP), (iii) logistic regression, (iv) a boost-
ing algorithm (AdaBoost), and (v) Bayesian networks.
All of the benchmark models are fixed-effect models
as these are the most common in the readmission pre-
diction literature. This permits a comparison of our
novel method of employing random effects to read-
missions with the literature standard.
For Classification and Regression Tree (CART), we
used classification trees with pruning having a mini-
mum 10 observations for the parent nodes, at least
one observation per leaf, and the same observation
weight. For the Multilayer Perceptron Neural Net
(MLP) we used one hidden layer with learning rate of
0.3 and momentum of 0.2. For multinomial logistic
regression, we used 0.25 as the minimum significance
levels for the variables to remain in the model with
backward-forward model selection strategy. For
Boosting, we employed ADABOOST M1 PART
with Decision Stump classifier and weight thresh-
old = 100. For Bayesian Net, we consider a simple
Bayes estimator with a = 0.5 and a hill climbing
search algorithm.
The analysis was performed based on data from a
database of 2449 patients with 3108 admission/read-
mission records, and 15 demographic, socioeconomic
and patient health-related factors from a medical cen-
ter in Michigan over the years 2006–2011; and from
the SID database, though the analysis is presented
only for the Michigan hospital since the results and
insights were similar for the SID database. After
initial analysis, the following nine variables were
incorporated into the model (several factors were
eliminated): length of stay, gender, age, employment
status, insurance coverage level, profession/military
rank, ward(s) visited during inpatient stay, principal
diagnosis, and source of admission, that is, VA hospi-
tal, nursing home, home, non-VA hospital. We used
threefold cross validation for evaluating model per-
formance. That is, we divided the data into three sep-
arate datasets, one for training, one for validation,
and one for testing; repeating this procedure three
times. For each of the three repetitions of the threefold
cross validation, the data were randomly divided
with 60% for training, 20% for validation, and 20% for
testing.
We used an iterative optimization process (in our
case simulated annealing) to determine the optimal
value of the parameters for the weight function,
W2ðtÞ, based on the validation dataset. The objective
function of the simulated annealing algorithm was
taken as the MSE of the estimated probabilities of re-
admission for patients in the validation set using the
model parameterized in the training dataset vs. the
respective empirical probabilities in the validation
dataset. However, weighting parameters complicate
the likelihood function and consequently affect the
efficiency of the Bayesian updating procedure. To
deal with this, we employ a simple, yet effective tech-
nique of replicating admission/readmission records
based on their weights. For example, if there are two
records where one of them has weight two, while the
other is a regular record with weight one, we may
replicate the first one twice and leave the last one
without replication. The appropriate number of repli-
cations for each record in more complex scenarios can
be achieved by using the least common multiple of
the weights.
Figure 3 presents the MSE in predicting individual
patient’s probability of readmission based on applica-
tion to the testing data in the threefold cross-valida-
tion using approximately 620 records (20% of the total
3108 readmission records). Four time snapshots (30,
60, 180, and 360 days after discharge) were used to
Table 1 High-Level Procedure of the Readmission Prediction
Framework
Readmission prediction procedure
Input Readmission data, Prior dist. of hazard function
parameters and risk factor covariates
Output Posterior dist. of hazard function parameters
and risk factor covariates
Empirical time to readmission density function
Procedure 1. Set prior dist. for risk factors, frailties, and marginal
baseline hazard (e.g., double exponential,
multivariate normal, gamma)
2. Set W2 function parameters (record recency weight)
3. FOR patient i in dataset
4. Calculate the total weight of all records in
the database with respect to the last readmission
event of patient i
5. Replicate each record according to the least
common multiplier of the calculated weights
6. Find the posterior distribution of the model
parameters based on the weighted dataset
using Gibbs sampling
7. Calculate the posterior distribution of readmission/
no readmission based on optimal parameters
NEXT PATIENT
Return Posteriors

compare our model against its counterparts. It is clear
that our method significantly outperforms other lead-
ing prediction methods for all time horizon predic-
tions.
3. Stage 2: Tactical and Operational
Planning Model for Design and
Optimization of Post-Discharge
Follow-Up Planning
Recent studies have showed that targeted post-dis-
charge telephonic outreach, telemedicine follow-ups,
and dedicated discharge management teams can
reduce the number of readmissions (e.g., Cardozo
and Steinberg 2010, Melton et al. 2012, de Toledo
et al. 2006). Our clinical collaborator and medical col-
leagues have also expressed an interest in more effec-
tive monitoring approaches employing a variety of
methods including phone calls and office visits. In
this section, we design a tactical planning model that
integrates foreknowledge of patient readmission rates
from section 2 with an optimization model to deter-
mine (i) what days a follow-up organization (clinic,
telemedicine, etc.) should be open, (ii) how many
resources to allocate to each specialty/type of follow-
up, and (iii) when to schedule discharged patients for
a follow-up.
3.1. A Delay-Time Model of the Post-Discharge
Patient Health Condition
The accurate empirical forecast of patient readmission
timing from section 2 paves the way for improved
planning and scheduling of at risk patients to avoid
or mitigate readmissions before they occur. In parti-
cular, the forecasting models form the basis of a
delay-time model of patient readmission dynamics.
Delay-time models have been employed extensively
in machine maintenance literature (e.g., Keller 1974,
Luss 1976) but are not as common in health-care
applications. Let J be the set of patient types (e.g.,
bladder cancer, hip replacement, etc.), and pjðtÞ be the
empirical probability distribution developed in sec-
tion 2 representing the time from when a patient of
type j 2 J is discharged until they are readmitted. In
addition to pjðtÞ, we also consider how long the
patient had a detectable condition before getting read-
mitted to the hospital, and thus presented an opportu-
nity to detect the condition before it triggered an
emergency readmission. To do so, we consider that
after a type j patient is discharged they may develop a
condition at a later time that will eventually result in
readmission. Once they develop a condition, there is a
delay, which we call Dj $ FjðtÞ, from the time the
patient develops the condition to the time they get re-
admitted to the hospital. If a patient has an “inspec-
tion”—for example, follow-up office visit or phone
call—after they develop the condition but before they
are readmitted then the doctor can take steps to avoid
or mitigate the readmission. Figure 4 depicts the read-
mission dynamics at a high level.
Traditional delay-time models start with a distribu-
tion on the time that a condition develops and com-
bine it with the delay until that condition causes a
readmission. This study takes the reverse approach,
since time to readmission is directly observable from
the data, whereas the time when the patient develops
the condition is not. Thus, the methods in section 2
are capable of directly calculating the distribution on
the timing of a readmission and not the time when
the patient developed the condition. Thus, we directly
employ the time to readmission curve from the esti-
mation model and use the delay-time formulation to
calculate the time when the condition develops.
Let T be the length of the planning horizon (e.g.,
30 days after discharge) and let T ¼ f0; 1; 2; . . .; Tg
be the set of days in the planning horizon, where the
patient’s discharge day is 0 (i.e., zero days after dis-
charge). The probability that a patient has developed
a condition but not yet been readmitted by t1 days
after discharge (i.e., the condition is detectable) is
given by
RT
t1
pjðsÞ½1 À Fjðs À t1ÞŠds. The other impor-
tant dynamic in the readmission system is that the
number of discharges and health-care resource capac-
ities vary by day of week, leading to a weekly repeat-
ing cycle of d0 ¼ 0; . . .; 6 for days of the week in
the discharge cycle and d1 ¼ 0; . . .; 6 for days of the
week in the resource capacity cycle. For the rest of the
0.00
0.05
0.10
0.15
0.20
0.25
30 days 60 days 180 days 360 days
MSE
Days after discharge
Our Approach CART MLP
Logistic Regression Boosting Bayes Net
Figure 3 Mean Squared Error (MSE) of the Comparing Methods
Hospital Home
Home + Readmission Causing
Condition
Figure 4 Post-Discharge Patient Flow with Readmissions

study, we will use ti to refer to the patient’s 30 day
planning horizon after discharge, and di to refer to the
cycle of the health-care system (discharges or capac-
ity). Letting Y
j
d0
be the random variable for the num-
ber of type j patients who are discharged on day d0 of
the planning horizon, we can show that the number
of patients that have developed a condition, but have
not yet been readmitted (I) by day t1, N
j
I;d0
ðt1Þ, has a
Binomial mixture distribution with mean
E½N
j
I;d0
ðt1ÞŠ ¼ E½Y
j
d0
Š
Z T
t1
pjðsÞð1 À Fjðs À t1ÞÞds: ð7Þ
It is easy to show that, for discharges that follow a
non-homogeneous Poisson process, this distribution
follows a Poisson distribution.
Now we can calculate the number of readmissions
of type j patients who were discharged on day d0 that
are averted or caught early (A) by scheduling a fol-
low-up on day t1, N
j
A;d0
ðt1Þ. This depends on the type
of inspection and how well it can detect a readmission
causing condition. The set of inspection types is given
by R. In our case, we consider one type of perfect and
imperfect inspection corresponding to an office visit
or a phone call, respectively (i.e., R ¼ fper; impg).
These are two options that are currently in use in an
ad hoc manner and are being considered for imple-
mentation by our clinical co-author. Other options
include telemedicine and home visits.
Let rk be the probability of detecting a condition
using follow-up method k 2 R (for perfect inspection
rper ¼ 1 and imperfect rimp 1). To calculate the prob-
ability of detecting a condition at the n + 1st follow-
up, first note that it is only necessary to know the
history of follow-ups following the last perfect
inspection. This is because a perfect inspection “wipes
the slate clean” by detecting all conditions that have
occurred and not yet caused a readmission prior to
that inspection. Thus, let t0 be the time of the most
recent perfect inspection before time t with t0 ¼ 0 if
there has not yet been a perfect inspection. If we let
PðT Þ be the power set of T minus the null set (since
the null set lacks meaning in our context), then we
can represent the history of imperfect inspections for
a patient that had their last perfect inspection at time
t0 as s 2 PðT Þ ¼ ft0; t1; t2; . . .; tn : titiþ1; i ¼ 0; . . .;
n À 1g. We define the set of feasible actions that can
follow a particular history, s, as
AðsÞ ¼ fðper; tÞ : T ! t [ max
s2s
sg
[ fðimp; tÞ : T ! t [ max
s2s
sg [ fEndg
ð8Þ
which represents all possible timing/type combina-
tions for the next follow-up after the sequence of
follow-ups, s, where (imp,t) and (per,t) represent an
imperfect and perfect inspection, respectively, at
time t and “End” represents the action of doing no
more follow-ups for the remainder of the planning
horizon. For purposes of exposition, we allow for a
slight abuse of notation by defining €tðaÞ and €kðaÞ to
be the time of the inspection associated with action
a (e.g., tnþ1 from above) and the type of inspection
associated with action a (e.g., perfect or imperfect),
respectively. Further, let €nðsÞ be the number of
imperfect inspections in history s and ~tiðsÞ be the
time of the ith imperfect follow-up in history s for
i ¼ 1; . . .; €nðsÞ and ~t0ðsÞ be the time of the most
recent perfect inspection follow-up. We use the con-
vention (with abuse of notation) that ~t0ðsÞ ¼ À1 if
there has not yet been a perfect inspection. It is
important to note that ~t€nðsÞðsÞ represents the time of
the most recent follow-up in the follow-up history s,
as this particular follow-up will be referred to fre-
quently. The probability of detecting a condition in
a type j patient at the n + 1st follow-up, given his-
tory s and action a 2 AðsÞ is given by
qjða;sÞ ¼ r€kðaÞ
Z T
s¼€tðaÞ
pjðsÞ½Fjðs À~t€nðsÞðsÞÞ
À Fjðs À€tðaÞÞŠds
þ
X€nðsÞÀ1
i¼0
Z T
s¼€tðaÞ
ð1 À rimpÞiþ1
pjðsÞ
½Fjðs À~t€nðsÞÀiÀ1ðsÞÞ À Fjðs À~t€nðsÞÀiðsÞÞŠds
!
ð9Þ
The multiplier outside the parenthesis accounts for
the detection rate of the type of follow-up indicated
by action a. Inside the parenthesis, the first term
accounts for all the conditions that have developed
since the last inspection (that occurred at time ~t€nðsÞðsÞ)
and hence could not be detected at prior inspections.
The second term accounts for the probability of
detecting all conditions that were (i) first detectable
during the imperfect inspection at time ~t€nðsÞ À iðsÞ, but
(ii) were not detected at that follow-up or any of the
subsequent i imperfect inspection follow-ups (hence
ð1 À rimpÞiþ1
) because of the failure of the imperfect
inspection, and (iii) were not readmitted before time
€tðaÞ. Thus, they would have a chance to be detected at
the inspection at time €tðaÞ indicated by action a. The
sum over i adds one term for each imperfect inspec-
tion that has occurred since the most recent perfect
inspection. The term where i ¼ €nðsÞ À 1 is slightly
different because ~t0ðsÞ represents the beginning of the
string of imperfect inspections, being either the most
recent perfect inspection (~t0ðsÞ [ 0) or when the
patient was discharged from the hospital if there has
not yet been a perfect inspection (~t0ðsÞ ¼ À1). In the

case where ~t0ðsÞ ¼ À1 (no perfect inspection yet),
then Fjðs À ~t0ðsÞÞ ¼ 1, which corresponds to the fact
that failures only begin to arrive after the patient is
discharged (since the discharge process is considered
a perfect inspection). Thus, 1 À Fjðs À ~t1ðsÞÞ repre-
sents conditions that were first detectable at the first
imperfect inspection and have not yet caused a read-
mission.
Equation 9 gives the probability of detecting a read-
mission causing condition for an individual patient.
We now link this detection probability to the decision
of how many appointment slots to reserve for follow-
ing up with a cohort of patients that are discharged
from the hospital. With a 2 AðsÞ being the timing and
type of the next follow-up we define this decision var-
iable as H
j;a
d0;s, which is the number of phone call slots
(if €kðaÞ ¼ imp) or office visit appointments (if
€kðaÞ ¼ per) to reserve on day €tðaÞ for type j 2 J
patients who were discharged on day d0 ¼ 0; . . .; 6
and have a follow-up history of s 2 PðT Þ. Each
follow-up has the opportunity of detecting a condi-
tion before it causes an emergency readmission. The
result may still be a readmission, but a less costly
one. Our assumption is that detecting the condition
before it becomes an emergency readmission averts
some or all of the cost of an emergency readmission
for that patient. This is what we mean by averting a
readmission. Letting 1 be the indicator function, and
recalling that Y
j
d0
is the number of patients of type
j discharged on day d0, the number of additional
readmissions averted by appointment allocation
H
j;a
d0;s is given by
N
H
j;a
d0;s
A;j ¼
X
Y
j
d0
^H
j;a
d0;s
‘¼1
1fCondition Detected at time
tnþ1 for Patient ‘g;
ð10Þ
where ^ represents the minimum operator. As
before, this also follows a Binomial mixture, which
can be seen by conditioning on Y
j
d0
and applying the
law of total probability. Further, it can be shown
that
E½N
H
j;a
d0;s
A;j Š ¼ E Y
j
d0
^ H
j;a
d0;s
h i
qjða; sÞ ð11Þ
Because the Θ’s capture both the follow-up day,
€tðaÞ, and the discharge day, d0, the optimization will
reveal not only the capacity to allocate to follow-ups
on each day d1 ¼ 0; . . .; 6 (by summing the decision
variable over f€tðaÞ : ð€tðaÞ þ d0Þ mod 7 ¼ d1g) but
also the optimal timing of each follow-up for a patient
of type j determined by the history s for all non-zero
Θ’s. The following sections will analyze the mix and
volume of follow-up types to determine how these
different methods (e.g., phone, office visit, etc.) can be
used most effectively in practice. The goal is (i) to
develop an optimization model that identifies optimal
placement of both types of follow-up, (ii) understand
the structure of the placement of these follow-ups to
provide heuristics or rules of thumb that could be
employed in designing post-discharge follow-up
plans even in the absence of optimization.
3.2. Optimal Design of a Post-Discharge
Monitoring Organization
While there is a rich literature on delay-time models
in machine maintenance, the optimization model we
develop adds new system-level decision-making
capability not previously considered by allowing the
optimization to consider not only the timing of
inspections but also how much capacity to reserve for
those inspections over a planning horizon. Adding to
the complexity, each day has a stochastic number of
discharges (potential patients to follow up with) with
a different distribution for each day of the week as
well as a different cost for scheduling follow-ups by
day of week. Finally, we develop a complex cyclosta-
tionary equilibrium model with parameters varying
over a seven-day horizon, but allowing patients to be
scheduled for inspection up to 30 days after their dis-
charge. Next, we present the notation followed by the
model formulation.
Let c
j;k
d1
be the cost per appointment of type k 2 R
for patient type j on day d1 ¼ 0; . . .; 6 (i.e., Sun-Sat).
uj;a is the usage requirement of follow-up resource
indicated by action a 2 s for a type j patient (e.g.,
some patients need 15-minute slots, whereas others
require 20- or 30-minute slots). bj is the average bene-
fit of averting a readmission by early detection for a
type j patient. Let ~Cd1;k be the maximum capacity of a
type k follow-up resource that can be reserved on day
d1. To capture the cyclically time-varying cost of
reserving capacity for follow-ups, we define a func-
tion that maps the follow-up day, t (on the scale of T ),
after a discharge on day d0 2 f0; . . .; 6g, to the day of
week the appointment would occur and correspond-
ing cost:
^cjða; d0Þ ¼ c
j;€kðaÞ
ðd0þ€tðaÞÞmod7
: ð12Þ
Finally, we need notation to capture the day of the
week that capacity is being reserved on to
account for the costs that differ by day. Let
Mðd0; d1Þ ¼ ft 2 T : ðt þ d0Þ mod 7 ¼ d1g be the
set of possible follow-up days of a patient dis-
charged on day d0 that map to the same day of the
week, d1. In this set, d0 is that day of the week of
discharge (e.g., d0 ¼ 0 represents a Sunday dis-
charge, . . .; d0 ¼ 6 represents a Saturday discharge)
and t þ d0 represents the (relative) date t days after

discharge. Therefore, ðd0 þ €tðaÞÞ mod 7 returns the
day of the week for that date that is t days after the
patient’s discharge. d1 2 f0; . . .; 6g represents a day
of the week (e.g., d1 ¼ 1 is Monday), so if
ðt þ d0Þ mod 7 ¼ d1, then this implies that t days
after discharge day d0 (e.g., Friday) is the d1 (e.g.,
Monday) day of the week. Thus, Mðd0; d1Þ is all the
days within the planning horizon (e.g., t = 1, . . ., 30)
that correspond to the day of week d1. For example,
if d0 ¼ 5 (i.e., Friday) and d1 ¼ 1 (i.e., Monday),
Mðd0; d1Þ ¼ f3; 10; 17; 24g as 3 days after Friday is
Monday, 10 days after Friday is also Monday, etc.
PROGRAM 1 provides an optimal design for staff-
ing and scheduling of a follow-up organization.
PROGRAM 1:
min
H
E
X
j2J
X6
d0¼0
X
s2PðT Þ
X
a2AðsÞ
^cjða; d0ÞH
j;a
d0;s Á uj;a À bjN
H
j;a
d0;s
A;j
2
4
3
5
ð13Þ
s.t.
~Cd1;k !
X
j2J
X6
d0¼0
X
s2PðT Þ
X
a2AðsÞ:€kðaÞ¼k;€tðaÞ2Mðd0;d1Þ
H
j;a
d0;s Á uj;a
8d1 2 f0; . . .; 6g; k 2 R
ð14Þ
X
a2AðsÞ
H
j;a
d0;s H
j;ðimp;~t€nðsÞðsÞÞ
d0;snfðimp;~t€nðsÞðsÞÞg
8j 2 J ; d0 ¼ 0; . . .; 6; s 2 PðsÞ : jsj [ 1
ð15Þ
X
s2PðsÞ:~t€nðsÞt1
H
j;ðper;t1Þ
d0;s !
X
k2R
XT
t2¼t1þ1
H
j;ðk;t2Þ
d0;ft1g
8j 2 J ; d0 ¼ 0; . . .; 6; t1 ¼ 2; . . .; T À 1
ð16Þ
The objective function minimizes the cost of stafﬁng
the follow-up clinic minus the cost of readmissions
averted. Equation 14 ensures that capacity con-
straints for each resource are respected. This is com-
plicated by the fact that we are considering a system
that functions on a weekly repeating cycle. Thus, we
have different costs and discharge distributions for
each day of the week, but the pattern repeats each
week. However, patient health status does not
evolve on a weekly repeating cycle but instead plays
out over a longer (non-repeating) horizon, which we
denote by T and can be thought of as the 30-day re-
admission window for example. In order to ensure
that enough capacity is allocated on a given day of
the week, take Monday, for example, we must
account for patients that are scheduled the ﬁrst
Monday after their discharge, as well as those that
will be scheduled the second Monday after their dis-
charge, and the third Monday after their discharge
and so forth. This is the reason for choosing the fol-
low-up day such that ð€tðaÞ þ d0Þ mod 7 is equal to
the day whose capacity is the focus of the constraint.
We also consider all patients who were discharged
on different days of the week (the sum over d0)
since the discharges vary by day of week in the
cyclostationary health-care system.
Finally, Equations 15 and 16 are critical constraints
linking sequential follow-ups together, ensuring that
the follow-up time sequence follows the correct pattern.
In particular, it should only be possible to schedule as
many follow-ups at time t1 in the sequence s [ ft1g as
the most follow-ups that were scheduled in any previ-
ous visit in s. This is because the probability of detection
for each patient is dependent on all the previous inspec-
tion times, and so to obtain the correct detection proba-
bility it is necessary that each patient being scheduled
in a particular time sequence has taken every visit in
the sequence. Equation 15 ensures that the number of
inspections scheduled following a history s (LHS of the
equation) is no more than the number of inspections
scheduled in the most recent imperfect inspection in s
(RHS of the equation). Equation 16 ensures the same for
perfect inspections. The reason we need two separate
constraints to capture this criterion is because a perfect
inspection “resets” the inspection history, and thus
must be treated differently.
The stochastic optimization model in PROGRAM 1
suffers from an extremely large state space, action
space, and number of constraints. Both the number of
decision variables and number of constraints in Equa-
tion 16 are exponential in the length of the planning
horizon because the probability of detection at each
visit depends on the history of all imperfect inspec-
tions since the last perfect inspection. Further, the fact
that the number of discharges on a given day can be
stochastic makes E½N
H
j;a
d0;s
A;j Š a non-linear function of the
decision variable, H
j;a
d0;s.
3.2.1. Computational Challenges. It is easy to see
that the number of constraints represented by Equa-
tion 15 is exponential because of PðT Þ. Proposition 1
shows that the number of decision variables in PRO-
GRAM 1 is also exponential in the length of the plan-
ning horizon.
PROPOSITION 1. Let T be the length of the planning hori-
zon. The number of decision variables in PROGRAM 1
is jJ jjRj Á 7ð2Tþ1
À ðT þ 2ÞÞ.
PROOF. At each time point, t, in the planning hori-
zon there are two choices: schedule a perfect inspec-
tion or an imperfect inspection. The outcome of the

choice depends on the history of imperfect inspec-
tions and the time of the most recent perfect inspec-
tion prior to t. If the last perfect inspection were at
time t À 1, there is only one possible history (i.e.,
20
) that has an inspection at time t. If the last perfect
was at time t À 2, there are two possible (i.e., 21
)
histories per, imp or per, None. If the last perfect
inspection was at time s then there are 2t À s À 1
possi-
ble histories. If there was no previous perfect inspec-
tion, there are 2tÀ1
histories. Thus, at time t, there
are a total of
Pt À 1
s ¼ 0 2s
¼ 2t
À 1 possible histories
and there are jRj possible actions at time t (in
our case jRj ¼ 2: perfect or imperfect inspection).
Thus, there are jRjð2t
À 1Þ decision variables at
time t. Summing this over all time points in the plan-
ning horizon yields jRj
PT
t¼0ð2t
À 1Þ ¼ jRjð2Tþ1
À
ðT þ 2Þ. Finally, this pattern repeats for each patient
type (of which there are jJ j) and days in the cost/
discharge cycle (d = 0, . . ., 6). h
For a 30-day planning horizon, the number of deci-
sion variables significantly exceeds commercial solver
limits. To handle the problem size, we introduce a
new, network flow-based method for transforming
the stochastic optimization problem into a tractable
linear deterministic one in the next section. We show
that this transformed problem is actually a weakly
coupled network flow problem (defined in section
3.3), which can be decomposed into several indepen-
dent networks with a few linking constraints. Not
only does the network flow formulation allow for
much faster solution approaches, the network flow
structure inherently captures constraints of Equations
15 and 16, leaving only a small number of linking con-
straints (Equation 14) that weakly couple otherwise
independent networks for each patient type and day
of week.
Finally, by applying pruning methods to our
decomposed network models, we are able to solve
even large problems very effectively. This allows us
to design complete patient follow-up monitoring
schedules as well as staffing plans for a realistically
sized follow-up organization. The development of
such techniques has the potential to impact many
areas by providing a methodology for solving multi-
dimensional, large state space stochastic optimization
problems that are common to the health-care domain
and elsewhere.
3.3. Network Flow-Based Transformation of the
Stochastic Optimization Model
Methods in the literature have been developed for
transforming these stochastic queueing network
problems into deterministic ones that admit tractable
optimization methods. Unfortunately, prior transfor-
mations in this vein of literature (such as Helm and
Van Oyen 2014, Helm et al. 2013, and Deglise-
Hawkinson et al. 2013) fail when applied to solve
reasonably sized post-discharge monitoring problems.
After applying previous transformation methods,
we could solve problems in a reasonable time for a
15-day planning horizon, as shown in Figure 6,
whereas our goal is to develop an optimal schedule to
reduce 30-day readmissions.
In this section, we show that our problem has a spe-
cial structure that allows us to decompose our prob-
lem into a set of independent network flow models
with a small number of linking side constraints. Simi-
lar to the naming convention of Gocgun and Ghate
(2012), we call this a weakly coupled network flow
problem. We develop methods to solve large problem
sizes that are intractable for general linear program-
ming representations of the readmission problem that
do not exploit the weakly coupled network structure.
Our particular model of post-discharge monitoring
exploits the fact that we can decompose the monitor-
ing problem along days of the repeating discharge
and follow-up cost cycle (e.g., days of the week),
along patient types, and along probabilistic sample
paths in the case of stochastic discharges. If there are
n patient types and a maximum of m possible dis-
charges on any given day then we would solve 7ÁmÁn
smaller networks (each of which typically solves in
seconds) instead of one large network (that is too
large to input into commercial solvers). Using
Lagrangian relaxation on the linking constraints, we
can decouple these 7ÁmÁn networks and solve them in
parallel, allowing for much larger problems to be
solved. In fact, we are able to solve the 30-day horizon
problem relatively quickly, even though the number
of potential decisions is extremely large.
We first present the network model with determin-
istic discharges and then extend the model to incorpo-
rate stochastic discharges. To do so, we show that the
non-linear stochastic objective, the expected number
of readmissions averted for any given capacity limit
(i.e., E Y
j
d0
^ H
j;a
d0;s
h i
from Equation 11), can be calcu-
lated exactly using a new stochastic branching
method that we develop. This method maintains the
weakly coupled network structure and allows us to
decompose the problem and tractably solve a set of
smaller network flows with side constraints. A further
convenient feature of our stochastic branching
method is that the number of “stochastic” branches
taken determines the optimal amount of capacity to
reserve on each day.
3.3.1. Deterministic Discharges. We begin by
describing the individual decoupled networks, one
for each discharge day and patient type, which
are the building blocks of the full, weakly coupled

network formulation. A truncated example of a de-
coupled network is depicted in Figure 5.
In this figure, Pt and It represent a perfect or imper-
fect inspection, respectively, on day t after discharge.
Each node in the network represents reserving capac-
ity for a particular inspection on a particular day,
given a specified history of inspections up to that
point. For example, node P2I4I7 represents scheduling
an imperfect inspection on the 7th day after dis-
charge, given that the most recent perfect inspection
was scheduled on day 2 and there was also an imper-
fect inspection scheduled on day 4. The amount of
flow that passes through a node represents the
amount of capacity to reserve on that day for that type
of inspection.
Each arc adds a new inspection (time and type) to a
previous sequence of inspections (the node the arc
originates from), by connecting the originating node
with a node that adds the new inspection to the previ-
ous sequence (if the inspection is imperfect) or starts a
new sequence (if the inspection is perfect). We denote
this concatenation/revision by the operator ⊕. Node
s represents inspection history s 2 PðT Þ up to time
~t€nðsÞðsÞ. Each arc from node s represents an action,
a 2 AðsÞ. Hence, from node s, arcs connect to all
future times, t [ ~t€nðsÞðsÞ, and all future inspection
sequences that can occur starting from the sequence s.
For patient type j, the cost of each arc is the cost of
having an inspection on day t (if the arc enters a node
associated with inspection type k on day t) minus the
expected reward of averting a readmission. The
expected reward that makes up part of the arc
cost, E½bjN
H
j;a
d0;s
A;j Š calculated from Equation 11, depends
on the history of inspections since the most recent per-
fect inspection (the node the arc originates from), and
the timing and type of the new inspection (the node
the arc terminates at). Thus, arc (s,s⊕a) for a type
j patient who was discharged on day d0 would have
cost
d
j;d0
s;sÈa ¼ ^cjða; d0Þ À bjqjða; sÞ; ð17Þ
where ^cjða; d0Þ is given by Equation 12 and qjða; sÞ is
given by Equation 9. We now present the weakly
coupled network flow formulation, where a small
number of side constraints couple the otherwise
independent networks described above. Let x
j;d0
iz be
the decision variable for the number of type j
patients on day d0 moving from inspection history i
to inspection history z = i⊕a for some action a.
Because the arc cost depends on all three parame-
ters—patient type, inspection history, day of dis-
charge—this actually represents an arc from node
ðj; d0; iÞ to node ðj; d0; zÞ. We differentiate arcs for
each patient type and discharge day in our notation
to highlight how the problem is weakly coupled and
how to decompose the full problem into smaller
subproblems. b
j;d0
s is the net traffic demand for node
ðj; d0; sÞ, where b
j;d0
n is equal to the number of dis-
charges at the source and sink nodes for patient
type j and discharge day d0 and is zero for all other
nodes. With |S| being the cardinality of set S, for
jJ j ¼ 2 and a discharge day cycle of length 7, there
are 7Á2 = 14 source nodes and sink nodes. Finally,
we define the set of inspection history transitions as
A ¼ fðs; s È aÞ : s 2 PðT Þ; a 2 AðsÞg.
Network Flow Formulation of PROGRAM 1
min
x
X
j2J
X6
d0¼0
X
ði;zÞ2A
x
j;d0
iz d
j;d0
iz ð18Þ
s.t.
X
ði;nÞ2A
x
j;d0
in À
X
ðn;zÞ2A
x
j;d0
nz ¼ b
j;d0
n
8n 2 PðT Þ; j 2 J ; d0 ¼ 0; . . .; 6
ð19Þ
X
j2J
X6
d0¼0
X
t2Mðd0;d1Þ
X
fs2PðT Þ:~t€nðsÞðsÞtg
x
j;d0
s;ftg
~Cd1;per
8d1 ¼ 0; . . .; 6
ð20Þ
X
j2J
X6
d0¼0
X
t2Mðd0;d1Þ
X
fs2PðT Þ:~t€nðsÞðsÞtg
x
j;d0
s;s[ftg
~Cd1;imp
8d1 ¼ 0; . . .; 6
ð21Þ
0 x
j;d0
iz UB
j
iz 8ði; zÞ 2 A; j 2 J ; d0 À 1; . . .; 6
I1
P2
I1I2
P1
I2
P1I2
Source
P3 I2I3 I3P2I3P1I2I3 I1I2I3P1I3 I1I3
End
Day 1
Day 2
Day 3
Days after
discharge
Figure 5 Decoupled Network Formulation Representing PROGRAM 1 for One Discharge Day and Patient Type and Three Follow-Up Days. Several
copies of this network will be solved in parallel with linking constraints moved to the objective by taking the Lagrangian

Equation 18 is the objective, which is a min-cost
flow. Equation 19 is the flow conservation con-
straint. Equation 20 is the capacity constraint for
perfect inspections. Every time there is a perfect
inspection at time t, it wipes the information set
clean except for the perfect inspection at time t.
Hence, the arc x
j;d0
s;ftg represents a perfect inspection
at time t given a history of s. t is chosen so that it
falls on day of the week d1 (see the sum over
t 2 Mðd0; d1Þ. To capture every possible perfect
inspection at time t, we also sum over all possible
imperfect inspection histories in which all the
inspections occur before t, given by fs 2 PðT Þ :
~t€nðsÞðsÞ tg, and recalling that ~t€nðsÞðsÞ is the most
recent inspection (largest inspection time) in s.
Finally, there is a sum over all possible discharge
days, d0, since patients discharged on any of the
discharge days in the cycle may be scheduled for
day d1 and use up some of the capacity. Equation
21 accomplishes the same purpose except for
imperfect inspections. All components are the
same as Equation 20, except in the index of the
decision variable, the information set is changed by
appending the new inspection at time t to the for-
mer history s; hence the arc x
j;d0
s;s[ftg goes from
inspection history s to inspection history s∪{t}
instead of erasing the old history as with the
perfect inspection.
3.3.1.1. Side Constraint Elimination and Network
Decomposition: Next, we discuss an efficient solution
method that is capable of solving large-scale
instances of the weakly coupled Network Flow
Formulation of PROGRAM 1. Inspection of the
problem structure reveals that the constraint set is
block diagonal, with the blocks linked together by a
relatively small number of the system capacity con-
straints defined in Equation 20 and 21. Further, each
block is an independent set of network flow con-
straints. Fortunately, there are only a few of these
side constraints relative to the number of constraints
that fit the network flow structure. By taking the
Lagrangian relaxation of these constraints it is possi-
ble to iteratively solve pure network problems and
use the subgradient method until the algorithm con-
verges to an acceptable tolerance (see Fisher 1985,
Geoffrion 1974). Further, Equations 20 and 21 are
the only constraints that link patient types and dis-
charge days to one another because of the sum over
j 2 J and d0 ¼ 0; . . .; 6. Once these constraints are
taken into the objective function, the full network
problem can now be decomposed into subproblems
with one pure network subproblem for each patient
type and discharge day. The subproblem looks the
same as the original except with a fixed j and d0 in
Equation 19 and in the objective as well as the
Lagrangian of the side constraints.
By taking advantage of this problem structure, we
are able to solve large-scale problems to optimality in
most cases, and near optimality in the remaining few
cases. Specifically, we relax the capacity constraints
Equations 20 and 21 into the objective function with a
corresponding vector υ of Lagrange multipliers. The
vector υ has jRj Á 7 elements (where in our case
jRj ¼ 2, corresponding to perfect and imperfect
inspection types), one multiplier per capacity con-
straint in Equations 20 and 21. The resulting relaxed
problem then further decomposes into 7 Á jJ j inde-
pendent network flow subproblems, one for each
patient type and discharge day of week, that can be
solved very quickly using subgradient optimization
(see Held et al. 1974) to search for the optimal multi-
pliers.
3.3.2. Stochastic Discharges Network Model. In
this section, we extend the deterministic model of
section 3.3.1 to consider the case of a stochastic number
of patients being discharged from the hospital. Unfor-
tunately, the presence of stochastic discharges not only
destroys the network structure, but it also eliminates
the linearity of the model. From the network stand-
point, with deterministic discharges the number of
patients scheduled was the same as the number
discharged. With stochastic discharges, you must pay
for the capacity reserved but you only get benefit for
the number of patients that actually get scheduled.
Thus, the concept of arcs from the deterministic model
has no meaning in the stochastic case because flow
means something different for cost (amount of slots
reserved) and for benefit (number of patients seen
times benefit). The capacity reservation remains linear
in amount of appointment slots reserved but from
Equation 11 the benefit from capacity reserved is
E Y
j
d0
^ x
j;d0
s;sÈa
h i
qjða; sÞ ¼ qjða; sÞ
Xx
j;d0
s;sÈa
n¼0
PðY
j
d0
! nÞ; ð22Þ
which is no longer linear, nor even convex. To over-
come this challenge and restore both linearity and
the network structure, we employ a stochastic
branching method we call sample path decomposition.
3.3.2.1. Sample Path Decomposition: To incorporate
stochastic discharges, we further decompose the net-
work along sample path realizations and then send
exactly one unit of flow through each subproblem.
Let xn be the realization of the discharge random var-
iable Y
j
d0
that corresponds to the number of discharges
being larger than or equal to n. Thus,
PðxnÞ ¼ PðY
j
d0
! nÞ. If the maximum possible num-
ber of discharges is M, then we decompose each
deterministic subproblem (with a source supply of
b
j;d0
0 ! 1) into M subproblems with a source supply of

b
j;d0;xn
0 ¼ 1 for n = 1, . . ., M. Then, the cost of each arc
(similar to Equation 17) is given by
d
j;d0;xn
s;sÈa ¼ ^cjða; d0Þ À bjPðxnÞqjða; sÞ; ð23Þ
Now x
j;d0;xn
iz is the decision variable indicating
whether the nth patient of type j discharged on day d0
moves from inspection history i to inspection history
z = i⊕a for some action a. The objective becomes
minx
P
j2J
P6
d0¼0
PM
n¼1
P
ði;zÞ2A x
j;d0;xn
iz d
j;d0;xn
iz . Note this
objective calculates exactly the desired objective of
Equation 22. The flow balance constraints remain the
same except for an addition of the xn index to Equa-
tion 19. Recall that the sample path decomposition
now defines flows as actual patients arriving to be
seen rather than slots reserved, but it is necessary to
pay for all slots reserved, whether or not patients are
seen. However, in each subproblem the full price of
capacity is charged for each unit of flow sent across
the arc, but the benefit is discounted by the probabil-
ity of the patient showing up, PðxnÞ.
The final challenge is how to deal with capacity
constraints. Again the flows represent patients seen,
but we must reconcile this with the fact that capacity
constraints refer to appointments reserved. This is
handled as follows. For the nth subproblem (xn), the
solution of the network flow subproblem will send
zero flow through a resource on a particular day
(even if positive flow was sent at a xm for m n),
indicating that the probability that the nth patient will
show up is not worth the cost of reserving an appoint-
ment for that patient. Thus, for each day of the plan-
ning horizon (d1 ¼ 0; . . .; 6), let m be the last network
with positive flow on that day. Then, m is the amount
of capacity that should be reserved. Summing the
benefit from each network matches the definition of
the objective in Equation 22, in which the LHS is the
the expected benefit from reserving x
j;d0
s;sÈa capacity
and the RHS is exactly the calculation that we get by
summing the non-zero flows from the xm subprob-
lems. The costs also match, since each flow is charged
the full cost of the capacity reserved. Thus, the capac-
ity calculation in the Lagrangian relaxation will be
correct as well.
After applying the sample path decomposition, it is
now possible to regain linearity and the pure network
form while only increasing solution times linearly in
M (the maximum possible discharges), although
using parallel computing solution times may not
increase at all. This is true because it is only necessary
to solve M additional subproblems (xi for
i = 1, . . ., M) for each original subproblem from the
deterministic optimization and these subproblems
can be solved in parallel similar to the previously
described decompositions. Using the pruning meth-
ods developed in the next section, however, it will
become clear that in most cases it will not be
necessary to solve anywhere near M subproblems to
incorporate stochasticity as PðxiÞ will quickly become
so small that all branches can be pruned, thereby
eliminating the need to solve any xj for j i.
3.3.3. Pruning Arcs/Nodes
Even with the weakly coupled network formulation,
the size of the problem still hinders us from solving a
full 30-day planning horizon as the number of arcs
and nodes grows too large for commercial solvers.
Computational results supporting this are presented
in section 3.3.4. Fortunately, the special structure of
our problem enables us to prune a large number of
arcs and nodes from the network without impacting
the optimal solution. The following theorem formal-
izes this notion by guaranteeing that certain arcs will
never be taken in the optimal solution.
THEOREM 1. In the Network Flow Formulation of PRO-
GRAM 1 (stochastic and deterministic), an arc with
positive cost will have zero flow in the optimal solution
to the network flow.
PROOF. We prove the result by showing that any
solution with a positive cost arc can be improved by
redirecting the flow away from the arc. Thus, in any
optimal solution no flow will be placed along posi-
tive cost arcs. Consider such a solution where arcs
ðs1; s2Þ and ðs2; s3Þ both have non-zero flow and the
costs of arc ðs1; s2Þ is c12 [ 0 and the cost of arc
ðs2; s3Þ is c23 0. If we send flow on arc ðs1; s2Þ and
ðs2; s3Þ along the arc ðs1; s3Þ instead we gain c1;2
because we are no longer sending flow along the
positive arc. Further, the cost of arc ðs1; s3Þ is better
than the cost of ðs2; s3Þ (i.e., c13 c23) because
removing the inspection in between the inspection
at s1 and the one at s3 only increases the value of
the inspection at s3, since it is less likely that a pre-
vious inspection has caught the potential readmis-
sion. This can be seen directly from Equation 9.
Also from Equation 9, the costs of all subsequent
inspections after node s3 are also either improved or
not impacted because having one fewer inspection
in the history of inspections does not decrease the
value of all subsequent inspections. Thus, redirect-
ing the flow on the positive valued arc results in a
lower overall cost flow. h
Not only does Theorem 1 enable the pruning of a
large portion of the arcs and nodes for any given net-
work, it also leads to the result that the optimal sche-
dule of a longer planning horizon can often be
obtained by solving a network for a much shorter
planning horizon. The following corollary states this
formally.

COROLLARY 1. Let VÃ
j;d0
ðtÞ (deterministic) or VÃ
j;d0;xðtÞ
(stochastic) be the optimal value of subproblem of the
network flow formulation of Program 1 for type j
patients discharged on day d0 (with sample path x for
stochastic) under a planning horizon of length t. S be
the time s.t. max
t [ S
fbjqjððper; tÞ; f0gÞ À ^cjððper; tÞ; d0Þg 0
(deterministic) or max
t [ S
fbjPðxÞqjððper; tÞ; f0gÞ À
^cjððper; tÞ; d0Þg 0 (stochastic). Then VÃ
j;d0
ðtÞ ¼ VÃ
j;d0
ðSÞ
(similarly VÃ
j;d0;xðtÞ ¼ VÃ
j;d0;xðSÞ) ∀t S. Further, the
optimal schedules will be the same for all t S.
PROOF. We show the result for the deterministic
case because the stochastic case is identical except
for a different reward term. It can be verified from
Equation 9 that the largest possible marginal benefit
that can be gained from scheduling an inspection on
day t is if the inspection scheduled is a perfect
inspection and is the first inspection of a patient’s
monitoring regime. To see this let ~t€nðsÞ þ 1 ¼ t and
note that for all s 6¼ {0} (i.e., schedules in which t is
not the first inspection of the monitoring regime)
qjððper; tÞ; f0gÞ ¼
Z 1
s¼t
pjðsÞ½1 À Fjðs À tÞŠds
À ^cjððper; tÞ; d0Þ
!
Z T
s¼€tðaÞ
pjðsÞ
X€nðsÞ
i¼0
½Fjðs À~t€nðsÞÀiðsÞÞ
À Fjðs À~t€nðsÞÀiþ1ðsÞÞŠds ! qjða; sÞ
The first inequality follows because
P€nðsÞ
i¼1 ½FjðsÀ
~t€nðsÞÀiðsÞÞ À Fjðs À~t€nðsÞÀiþ1ðsÞÞŠ þ Fjðs À~t€nðsÞðsÞÞ 1. The
second inequality follows because rimp
1. Thus
qjððper;tÞ;f0gÞ is the largest possible benefit of
having an inspection on day t. If this benefit is
less than the cost of scheduling an inspection on
day t for all t S, then all the arcs for days t S
will have positive cost and thus will be pruned by
Theorem 1. Thus, the network for horizon length
S will be identical to the pruned network for hori-
zon length t S and hence will have the same
optimal solution. Corollary 1 can be used to easily
compute off-line the finite horizon length needed to
achieve an infinite horizon optimal and thus greatly
reduce computation times. In section 3.3.4, we show
that Theorem 1 and Corollary 1 have a profound
impact on solution times, enabling us to solve
problems that were intractable in the original for-
mulation and even in the network formulation of
PROGRAM 1. h
3.3.4. Computational Results. This section dis-
cusses solution times for the various approaches to
solve the optimization problem described above. For
the original PROGRAM 1, the optimization fails to
solve except for instances with small planning hori-
zons (small T). Figure 6 demonstrates the benefit of
the network transformation as well as pruning. These
computation times are for one iteration of the subgra-
dient optimization for the Lagrangian relaxation of
the network flow formulation of PROGRAM 1 solved
on a computer with an Intel i5-3230M @ 2.6GHz pro-
cessor with 8GBs of RAM. The total times are linear in
the number of iterations of the subgradient optimiza-
tion, which are still typically 30 seconds for a 30-day
planning horizon.
Figure 6a demonstrates the impact of pruning on
solution times. Network subproblem decomposition
and pruning enabled by Theorem 1 and Corollary 1
can solve large problem instances that quickly
become intractable for the standard formulation and
even the full network formulation. It also shows
that pruning is effective regardless of the skewness
of the time to readmission density (pjðtÞ), demon-
strated by the “Right Justified” curve, in which the
peak of pjðtÞ was shifted to the right. Figure 6b
demonstrates the insight provided by Corollary 1,
as it is clear that once the model exceeds a fixed
planning horizon length the number of solution arcs
stops growing. This is because all future days (e.g.,
beyond 10 days for Pruned Network 5K Benefit) do
not have enough benefit and therefore get pruned.
The “Right Justified” curve also shows that prun-
ing can happen at the beginning of the planning
horizon as well as the end, as all arcs are pruned
before day 10.
4. Case Study: Numerical Analysis and
Insights
Given the increasing attention on hospital readmis-
sions by both policymakers and health-care adminis-
trators, a primary goal of this study is to provide
hospitals with an effective data-driven method for
reducing readmissions. Here, we present results illus-
trating the effectiveness of our approach at achieving
this goal. We focus our discussion on the size of the
reductions possible, the value of being able to predict
patient readmission risk profiles, and insights from
the patient follow-up schedules.
In this section, we employ the empirical prediction
model developed in section 2 to generate inputs to the
optimization model. For the patient readmission
curves, we randomly selected 10% of patients from
the full dataset described in that section. We then run
a series of experiments to generate insights into the
impact of optimal post-discharge monitoring, risk
profiling, stochasticity of discharges, time-varying

capacity limits and costs, and the benefit of averting a
readmission on resulting monitoring schedules, staff-
ing plans, and effectiveness of the post-discharge
monitoring system. We find that risk profiling (e.g.,
categorizing patients as high, medium, and low risk
for readmission) has a major positive impact on the
effectiveness of post-discharge monitoring.
The prediction model provides both the time to re-
admission density function, pjðtÞ, as well as a risk pro-
filing approach for separating patients into different
risk categories for readmission. This empirical density
function and classification scheme are then used to
build an optimization model for post-discharge moni-
toring. The discharge pattern, X
j
d0
, is based on
1.5 months of discharge data from our partner hospi-
tal. We based the delay-time function on a survey of
five surgeons in which we asked them to estimate
how long a patient might stay at home with different
conditions (e.g., infection, dehydration, etc.) before
being readmitted to the hospital. The averages for
major causes of readmission are as follows: (Infec-
tious, 48 hours), (Metabolic, 48 hours), (Failure to
thrive, 72 hours), (Urinary, 48 hours), (Hematologic,
48 hours), (Cardiac, 24 hours), (Pulmonary, 48 hours),
(Gastrointestinal, 48 hours), (Neuro/psych/MSK/
Oto/Optho, 48 hours), (Vascular, 24 hours), (Wound
related/hematoma, 72 hours). We used these data to
fit a discrete delay-time function with a mean of
36 hours. The detection probability of a phone call
was placed at 40% based on discussions with our clin-
ical co-author, but was also varied in the sensitivity
analysis.
The costs for phone calls and office visits were
determined through discussions with our clinical co-
author and verified with Medicare reimbursement
structures. For example, a phone call from a nurse
practitioner is reimbursed at the rate of $25 for a call
between 11 and 20 minutes. The office visit is more
complex because diagnostic tests may need to be
ordered in addition to compensating the doctor for
their time. Our clinical co-author estimates a follow-
up office visit would cost anywhere between $100
and $500 per visit depending on how many diagnos-
tics were needed, but most likely closer to $100–$200.
We then varied the costs of these tests by day of
week to capture the fact that some days, such as
weekends, are undesirable for performing follow-
ups from the medical professional and patient stand-
point. The benefit of averting a readmission (bj) is
varied throughout the case study as a means of
providing sensitivity analysis and insight. Recall that
the parameter bj represents a weighted average of
those readmissions that are completely eliminated
and those readmissions whose length and/or sever-
ity are reduced by early detection of the readmission
triggering condition.
4.1. Numerical Analysis
4.1.1. Risk Profiling. To develop a risk profile, we
used the empirical prediction model from section 2 to
generate personalized time to readmission curves for
each patient in our dataset. Using K-means clustering
based on total probability of readmission within
30 days we separated the patients into high, medium,
and low risk patients. The groups had an average 30-
day readmission probability of 72% for high risk, 18%
for medium, and 4% for low. In our dataset, 20% of
patients were high risk, 24% were medium risk, and
56% were low risk.
Figure 7 reports the percent of readmission trigger-
ing conditions that were detected before causing an
emergency readmission by optimally scheduling
patient follow-ups. Figure 7 compares a schedule
based on optimizing three-patient risk profiles with a
schedule based on optimizing a single population
aggregate readmission curve. The bars in Figure 7 cor-
respond to the percent averted readmissions on the
left y-axis and the solid and dotted lines correspond
to the follow-up schedule costs on the right y-axis.
Our results show that we are able to reduce the num-
ber of readmissions by roughly 40–70%, depending
on the cost benefit per averted readmission. Notice
that the vast majority of averted readmissions are for
patients in the high risk category. While this behavior
0
500
1000
1500
2000
10 12 14 16 18 20 22 24
SolutionTime(sec)
Length of Follow-Up Horizon (T)
Full Network - 10k
Pruned Network - 5k
Pruned Network - 10k
Pruned: Right Justified, 10k
Original Program1
0.1
10
1000
100000
10000000
1E+09
0 5 10 15 20 25
#ofDecisionVariables
Length of Follow-Up Horizon (T)
(a) (b)
Figure 6 Comparing Pruned vs. Unpruned Networks at Different Benefit Levels for Averting Readmissions (bj ) and Different Planning Horizon
Lengths (T). The same legend is used for both (a) and (b)

is as one would expect, it is important to note that it is
not a simple matter to accurately predict readmission
risk profiles. The techniques and empirical results
from section 2 allow hospitals to predict risk profiles
for their discharged patients. The optimization model
in section 3 then takes advantage of these results to
reduce readmissions by targeting the right patients
at the right time for follow-up phone calls and office
visits.
From Figure 7 it is clear that risk profiling has a
significant impact on the effectiveness of a post-
discharge scheduling and staffing plan. The risk profiled
approach (the solid bars) averts a much higher per-
cent of readmissions—increasing from 9% to 39%
averted, that is, more than four times as many readmis-
sions averted at $2K benefit—when the benefit of an
averted readmission is lower by focusing on the high-
risk patients that provide the most benefit per inspec-
tion. When the benefit of an averted readmission
grows, the difference in readmissions averted shrinks
but the risk profiled plan (solid line) does so at signifi-
cantly lower cost—2/3 the cost—than the aggregate
plan (dashed line), again by using targeted follow-
ups and not wasting too much effort on low-risk
patients. Risk profiling is also more effective when
capacity is more tightly constrained (Figure 7a) than
when it is not (Figure 7b). This is noticeably illus-
trated when the benefit per averted readmission is
$10k or $8k. At both levels, the percentage of averted
readmissions gap between the three risk profiled bars
and the aggregate bar is larger with capacity con-
straints in place.
Also note how few readmissions are averted from
the low risk profile even at a benefit of $10k per read-
mission averted. It requires planning for 56 phone
calls a week to avert 0.2 readmissions. From this, we
could draw the conclusion that low risk patients need
not be planned for but instead should be contacted
only when extra time is available, which can smooth
the workload and fill in gaps left by no-shows or
inability to fill all slots due to stochastic discharges.
Our empirical prediction model could provide a
printout each day of the optimal low risk patients to
target and the order in which to call them if the
nurse/doctor has time in their schedule.
Figure 8 reports the number of perfect and imper-
fect follow-ups staffed by day of the week and by
patient risk profile at a benefit of $10K and $6K per
averted readmission. In both cases, no perfect follow-
ups are scheduled for low risk patients, only phone
calls. The majority of perfect follow-ups are scheduled
for Monday, Wednesday, and Friday. Monday and
Friday are higher because, in order to avoid the week-
end, the patients whose optimal follow-up time
would have been on a Saturday get pushed to Friday
and Sunday visits get pushed to Monday. Because the
time to readmission distribution is mostly concave,
the “next best” follow-up time is adjacent to the best
one. Wednesday is used because it is a cheap day and
spaced out from Monday and Friday, as scheduling
back to back perfect inspections is not very beneficial.
At $10K benefit, appointments are reserved for high
risk patients on every day of the week but Sunday
(the most expensive day), because scheduling an
inspection at the “right time” is very beneficial due to
the high likelihood that these patients will eventually
be readmitted. For the other groups, however, the
schedule sticks to the Monday, Wednesday, Friday
staffing schedule mentioned previously. At the lower
$6K benefit, even the high risk patients are mostly
staffed for on Monday, Wednesday, Friday and low
risk patients are not staffed for at all. Even with this
comparatively light staffing plan, we are still averting
around 65% of all possible readmissions because of
the focused effort enabled by our empirical risk profil-
ing scheme.
Figure 9 shows the uncapacitated follow-up sche-
dule for a high risk patient in contrast to medium risk
0%
10%
20%
30%
40%
50%
60%
70%
80%
10,000 8,000 6,000 4,000 2,000
%AvertedReadmissions
Benefit per Averted Readmission
High Risk Med Risk Low Risk Aggregate
Risk Profiles Costs Aggregate Costs
10,000 8,000 6,000 4,000 2,000
Benefit per Averted Readmission
High Risk Med Risk Low Risk Aggregate
Risk Profile Costs Aggregate Costs
$0
$20,000
$10,000
$30,000
$40,000
Follow-UpScheduleCosts
(a) (b)
Figure 7 Comparing Effectiveness of Monitoring Plans with and without Risk Profiling. (a) Capacity Constrained (b) No Capacity Constraint

patients in Figure 10 for (a) uncapacitated optimal
and (b) capacitated optimal solutions. In all cases,
almost all follow-ups are scheduled within two weeks
of discharge. This is a significant finding because the
current standard after high-readmission rate surger-
ies is to follow-up 2–3 weeks after discharge. How-
ever, by this point most of the patients who would
need to be readmitted have already been readmitted.
Follow-ups in the optimal schedules typically begin
within 2 or 3 days of discharge, no matter the averted
readmission benefit (except $2000), always centered
around the peak of the readmission curve. In general,
it is best to place office visits (perfect inspections)
close to the peak of the readmission curve, with
phone calls before or after or both. As the benefit of
averting a readmission increases, the more the opti-
mal schedule will suggest calling patients further out
after discharge. For medium risk patients in Figure 10a
only one follow-up phone call and no office visits are
scheduled at the lowest benefit per averted readmis-
$2,000
$3,000
$4,000
$5,000
$6,000
$7,000
$8,000
$9,000
$10,000
$11,000
0 2 4 6 8 10 12 14 16 18 20
AvertedReadmissionBenefit
Horizon Days
$0
$50
$100
$150
$200
$250
$300
$350
$400
$450
0 2 4 6 8 10 12 14 16 18 20
PerfectVisitCost
Horizon Days
Perfect Visit Costs
Perfect
Imperfect
P(t) PDF
(a) (b)
Figure 10 Patient Follow-Up Schedule over a 30-Day Horizon (Tuesday Discharge) for a Medium Risk Patient Considering: (a) Uncapacitated (b)
Capacitated Optimizations. “3” is a phone call and “square” is an office visit. Dashed line gives the shape of the time to readmission
pdf pj ðtÞ and solid line is office visit cost by day of week.
0
10
20
30
40
50
Per Imp Per Imp Per Imp Per Imp Per Imp Per Imp Per Imp
Mon Tues Wed Thur Fri Sat Sun
AppointmentsScheduled
Low Risk
Med Risk
High Risk
0
10
20
30
40
Per Imp Per Imp Per Imp Per Imp Per Imp Per Imp Per Imp
Low Risk
Med Risk
High Risk
50
(a) (b)
Figure 8 Weekly Staffing Requirements for Follow-Ups Under Different Benefit Levels of Averting a Readmission
Figure 9 Post-Discharge Patient Follow-Up Schedule for High Risk Patients for Different Levels of Benefit of Averting a Readmission. “3” is a
phone call and “square” is an office visit. The solid line is the cost profile for office visits and the dashed line provides a structural repre-
sentation of the time to readmission density, pj ðtÞ.

sions, with more office visits gradually being added
as the benefit of averting a readmission increases. The
high risk patient is monitored closely over the first
17–20 days, while the medium risk patient generally
receives a few office visits during their peak-risk per-
iod shortly after discharge followed by a series of
phone calls.
To generate Figure 10b, we created capacity limits
based on discussions with our clinical co-author to
identify reasonable capacities for office visits and
phone calls. The effect of capacity limits by day of
week does two things to the schedule. First, it spreads
out the workload across the days of the week, with
less clustering around Monday, Wednesday, and
Friday. Secondly, as can be seen in Figure 10, capacity
limits can cause one type of follow-up to be replaced
with another. At the $4000 benefit level, for example,
the first two phone calls are replaced by an office visit
since phone calls were capacity constrained on that
particular day of the week. At $10,000, the fourth
office visit is replaced by two phone calls. This indi-
cates that it is very important to have a follow-up at
the right time and that, even though the phone call
has at most a 40% detection rate, adding several
phone calls at key times can be a good surrogate for
an office visit or vice versa.
4.1.2. Stochastic Discharges. The final component
we investigate is the impact of stochastic discharges.
We present the case where there are no capacity con-
straints to focus on the impact of stochasticity. Inter-
estingly, uncertainty regarding how many patients
will be discharged, and thus need follow-ups, does
not significantly change the timing of the follow-ups.
Adding uncertainty does, however, change the staff-
ing level across the week, encouraging the model to
move some capacity from office visits (more expen-
sive) to phone calls (less expensive) and from more
expensive days to cheaper days (Figure 11). The over-
all staffing level for expensive office visits is also
lower in the stochastic case because expensive
capacity is eliminated unless there is a high expecta-
tion the slot will be filled.
4.2. Sensitivity Analysis
The optimal monitoring schedules and the associated
number of readmissions averted are, of course, sub-
ject to the model inputs. Most figures in section 4.1
illustrate a sensitivity analysis on the key input, bene-
fit of an averted readmission. We did this because
each hospital and medical procedure will have a dif-
ferent value for the benefit of an averted readmission.
In this section, we perform a sensitivity analysis and
examine the impact of other important model inputs;
specifically the number of risk profiles, the detection
rate of an imperfect inspection, the structure of the
delay-time distribution, and how errors in the empiri-
cal prediction model propagate through the optimiza-
tion model.
First, we identify the impact of the number of risk
profile groups on the model solution by dividing the
patients in our dataset into two, three, four, and five
different risk profiles and comparing it to the aggre-
gated or one profile for the entire group. The results
are shown in Figure 12. The percentage of averted
readmissions is broken out in the bar chart for each
risk profile. This illustrates that almost all averted
readmissions are from the high risk profile. Please
note that the number of patients in each risk profile
changes as the number of profiles are increased. For
example, the number of patients in the Med-High risk
profile with only two profiles is much larger than the
number of patients in the Med-High risk profile when
there are five risk profiles.
Using two risk profiles, instead of leaving all
patients in one profile, improves averted readmis-
sions by 21% and reduces follow-up costs by 25.6%.
Moving from two to three risk profiles averts 6.1%
more readmissions. However, the cost of these fol-
low-ups increased 4.3%. Adding a fourth and fifth
risk profile provides minimal benefit in averted read-
missions (1–2%) and follow-up costs actually increase
slightly. Figure 12b illustrates why we choose three
risk profiles for our numerical analysis in section 4.1
The modeled solution value (Benefits À Total Cost)
increases from one to two to three risk profiles. How-
ever, after three risk profiles the solution value levels
0
20
40
60
80
$10,000 per Readmission Averted
Perfect
Imperfect
0
20
40
60
80
Perfect
Imperfect
0
20
40
60
80
Perfect
Imperfect
0
20
40
60
80
Perfect
Imperfect
0
20
40
60
80
Perfect
Imperfect
0
20
40
60
80
Perfect
Imperfect
0
20
40
60
80
Perfect
Imperfect
0
20
40
60
80
Perfect
Imperfect
(a) (b) St (c) (d)
Figure 11 Comparing a Staffing Plan Having Deterministic Discharges with a Plan Having Stochastic Discharges with the Same Mean for $10K
(a, b) and $6K (c, d) Benefit of an Averted Readmission

off, varying 1% thereafter. Thus, three risk profiles
provide the best balance of simplicity and efficacy.
All of the results in the sections above use a phone
call detection probability of 40%, however in Figure 13
we investigate the impact efficacy of a phone call by
varying the detection probability from 10% to 60%.
As expected, Figure 13b shows the optimization
model solution value (Benefits À Total Cost) increas-
ing as the detection probability improves. The results
in Figure 13a illustrate the general trend that costs
decrease and averted readmissions increase as the
detection probability of a phone call increases, though
this is not universally the case. Note that there is a
slight dip from 62.2% averted readmissions to 61.7%
when detection probability moves from 20% to 30%.
This occurs because of the discrete nature of timing
and type of follow-up. At 30% detection probability,
no phone calls are performed at all due to lack of
effectiveness, only office visits. At 30% detection
probability, the optimal solution starts to replace
some of the office visits with phone calls, leading to
an overall better solution (through lowered follow-up
staffing costs) but slightly lower detection rate.
Another key takeaway is that the objective value is
fairly robust to the detection probability, which
means that we can develop good solutions with a
wide variety of imperfect detection probabilities. Note
the averted readmission percentage only varies from
62% to 67% and the cost from $12K to $14.5K as the
detection percentage varies up to 60%.
We also analyzed the impact of changes in the
delay-time distribution on the optimal solution
results. Figure 14a shows the different delay-time
probability mass functions that were tested. These
delay-time variations are based on our survey of sur-
geons that indicated that delay times for typical read-
mission-causing conditions tend to range from 1 to
4 days, with the average delay time between 2 and
3 days. In this analysis, we examine the (i) sensitivity
to mean/skewness with curves #1 and #2, as these
two have the same variance but different means, and
(ii) sensitivity to variance with curves #3 and #4,
which have the same mean but different variances.
The results in Figure 14b behave as expected. When
the variance is held constant, the distribution with the
higher mean (Curve #1 ) averts 17% more readmis-
sions and has a lower cost than the distribution with
the lower mean (Curve #2 ). When the mean is held
constant, the distribution with the lower variance
(Curve #3 ) averts 7.6% more readmissions than the
$0
$3,000
$6,000
$9,000
$12,000
$15,000
$18,000
$21,000
$24,000
0%
10%
20%
30%
40%
50%
60%
70%
80%
Low
Med-Low
Med
Med-High
High
Costs
Averted
Readmissions
$0
$5,000
$10,000
$15,000
$20,000
$25,000
$30,000
Benefit–TotalCost
(a) (b)
Figure 12 Comparing Cost and Effectiveness of Averted Readmissions Across Differing Numbers of Risk Profile Groups
$0
$3,000
$6,000
$9,000
$12,000
$15,000
$18,000
$21,000
$24,000
0%
10%
20%
30%
40%
50%
60%
70%
80%
Low
Med
High
Costs
Averted
Readmissions
$0
$5,000
$10,000
$15,000
$20,000
$25,000
$30,000
$35,000
Benefit-TotalCost
(a) (b)
Figure 13 Comparing Cost and Effectiveness of Averted Readmissions Across Differing Imperfect Detection Levels

distribution with the larger variance (Curve #4 ). The
follow-up costs are slightly higher in Curve #3 with a
lower variance because more office visits were sched-
uled. This occurs since office visits on the optimal fol-
low-up day are more likely to find readmission
conditions with a low delay distribution variance.
Finally, we analyzed how errors in the empirical
prediction of the time to readmission pdf, pjðtÞ, would
impact the optimization model results. To test this,
we created a simulation to act in the same way a hos-
pital system would use our method in practice. Using
past data, a hospital system would determine the
optimal follow-up schedule for high, medium, and
low risk patients. There may be error in estimating
the readmission curve, pjðtÞ. Therefore, our simulation
used the optimal follow-up schedule and costs for the
high, medium, and low risk profiles with an averted
readmission benefit of $6000 and a 40% imperfect
detection rate. The simulation then randomly per-
turbed the readmission pdf, pjðtÞ, by an error factor
that was generated as a percentage in a positive or
negative direction where the percentage was drawn
from a uniform random variable. We then calculated
the readmissions averted based on the fixed follow-
up schedule and these randomly perturbed pdf
curves. The result of 1000 of these simulations can be
seen in Figure 15. This illustrates the robustness of the
optimization model to variation in the readmission
pdf, pjðtÞ, since the averted readmissions stay between
58% and 68% even as the allowable random error
approaches 100%.
4.3. Clinical Insights and Opportunities for Post-
Discharge Inspection Approaches to Minimize the
Burden of Readmissions
Using a combination of perfect (i.e., in-person office
visits) and imperfect inspections (i.e., automated
screening and live phone calls) to identify patients
who are susceptible to readmission is clinically feasi-
ble and made possible through the integration of
empirical prediction models (section 2), optimization
(section 3), and clinical knowledge. The statistical and
optimization tools developed in this study can offer
hospitals the ability to plan for special care at the time
of discharge with the confidence that they will have
the resources to support each individual’s follow-up
needs.
Implementing post-discharge inspection approaches
based on the models developed herein will reduce
the burden of readmissions in at least two important
ways leading to measurable decreases in readmissions,
subsequent length of stay and avoidable spending.
0
0.2
0.4
0.6
0.8
1
0 1 2 3 4 5
#1: E[D]=3.2, V[D]=0.56 #2: E[D]=1.8, V[D]=0.56
#3: E[D]=2.0, V[D]=0.1 #4: E[D]=2.0, V[D]=0.66
$0
$3,000
$6,000
$9,000
$12,000
$15,000
$18,000
$21,000
$24,000
0%
10%
20%
30%
40%
50%
60%
70%
80%
#1 #2 #3 #4
High Med Low Costs
Averted Readmissions
(a) (b)
Figure 14 Curves #1 and #2 are Used to Explore the Impact of the Mean and Skewness of D. Curves #3 and#4 are used to explore the impact of
variance of D, having the same mean but different variance. The mean (E[D]) and variance (V[D]) of each distribution are given in leg-
end (a)
52%
54%
56%
58%
60%
62%
64%
66%
68%
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Uniform Error Percentage (+/-) on P(t) Readmission Curve
Mean Averted Raw
90% High Confidence
90% Low Confidence
Figure 15 Impact of Errors in the Empirical Prediction of Time to Readmission pdf, pj ðtÞ, Propagating into the Optimization Model Results

First, by predicting resource needs for patient follow-
ups based on readmission risk, clinical deterioration
can be efficiently detected early through targeted office
visits and telephone calls to avoid readmission or at
least decrease its intensity. Many times, patients are
unaware of their clinical deterioration or simply take a
‘wait-and-see’ approach while their condition contin-
ues to worsen requiring more intense and expensive
care later on. For example, failing to identify an infec-
tion early so oral antibiotics can be given in the outpa-
tient setting will eventually result in an emergency
department visit, intravenous antibiotics, and hospital
readmission. Second, through identifying patients at
highest risk of readmission, hospitals would be able to
plan for readmissions and develop efficient systems to
deal with patients who return for inpatient care rather
than funneling them through expensive and resource
constrained emergency departments. This is especially
important since most current approaches to post-dis-
charge care are often ‘one-size fits all’ with customary
outpatient follow-up in two to three weeks regardless
of an individual patient’s readmission risk. As shown
above, our analyses suggest that tailored post-dis-
charge care optimized to each patient’s clinical sce-
nario and their health-care system has the potential to
significantly reduce readmission risk over time.
By taking into consideration the daily risk of clinical
deterioration and its detection, our models are able to
align appropriate inspection efforts to get patients the
supportive care they need in a timely fashion.
Although readily measured, the often cited 30-day re-
admission metric is much too blunt to direct large
scale clinical efforts to curtail readmissions. For these
reasons, the optimization tools developed herein have
the potential to direct guidelines for scheduling a
range of follow-up intensities based on risk factors, as
well as patient characteristics, and hospital character-
istics so as to reach the broadest group of health-care
organizations.
Finally, by using a State Inpatient Database (SID)
and a partner hospital for the validation portion of
this project, we were able to develop tools based on
readily available hospital discharge data thereby
increasing the generalizability of this work and its
subsequent implementation.
5. Conclusion
While reducing hospital readmissions is one of the
most pressing challenges for the US healthcare sys-
tem, few efforts harness the intersection of medicine,
statistics, and operations management to develop
innovative approaches to help solve the dilemma.
The comprehensive multi-methodology approach
combining empirical models to predict readmission
timing with optimization models to detect readmittable
conditions before they cause a readmission has the
potential to transform the way organizations
approach reducing the burden of readmissions on
patients and health-care organizations. Our approach
integrates machine learning and classical prediction
models along with transfer learning to generate accu-
rate and personalized predictions of time to readmis-
sion and enables classification of patients into risk
profiles. The results of the empirical model are then
integrated with methods to transform a large-scale
optimization based on stochastic delay-time models
into a weakly coupled network flow model with trac-
table subproblems that can be solved as independent
network flow models. In a case study based on data
from a partner hospital, we show that the empirical
model outperforms other leading techniques in terms
of predicting time to readmission and our optimiza-
tion model demonstrates the ability to avert 40–70% of
potential readmissions using simple and implementable
post-discharge follow-up schemes that simulta-
neously provide risk profiling, readmission timing
prediction, patient monitoring schedule design, and
staff planning. While this work focuses solely on
post-discharge patient management, future research
in readmissions may include integrating our work
with the impact of pre-discharge actions, such has
hospital length of stay, on readmissions.
Acknowledgments
The authors thank the senior editor and two anonymous
reviewers for their valuable comments and suggestions
which greatly improved the paper. Ted Skolarus gratefully
acknowledges grant support he received from VA HSRD
Career Development Award—2(CDA 12-171).
References
Cardozo, L., J. Steinberg. 2010. Telemedicine for recently dis-
charged older patients. Telemed. J. E-Health 16(1): 49–55.
Clayton, D. G. 1978. A model for association in bivariate life
tables and its application in epidemiological studies of famil-
ial tendency in chronic disease incidence. Biometrika 65(1):
141–151.
Deglise-Hawkinson, J., M. P. Van Oyen, B. Roessler. 2013. Sto-
chastic modeling and optimization in a decision support sys-
tem for phase 1 clinical trial management. Working paper.
Desai, M. M, B. D. Stauffer, H. Feringa, G. C. Schreiner. 2009. Sta-
tistical models and patient predictors of readmission for acute
myocardial infarction a systematic review. Circ. Cardiovasc.
Qual. Outcomes 2(5): 500–507.
Dunlay, S. M., M. M. Redfield, S. A. Weston, T. M. Therneau, K.
H. Long, N. D. Shah, V. L. Roger. 2009. Hospitalizations after
heart failure diagnosis: A community perspective. J. Am. Coll.
Cardiol. 54(18): 1695–1702.
Feudtner, C., J. E. Levin, R. Srivastava, D. M. Goodman, A. D. Slo-
nim, V. Sharma, S. S. Shah, S. Pati, C. Fargason, M. Hall.
2009. How well can hospital readmission be predicted in a
cohort of hospitalized children? a retrospective, multicenter
study. Pediatrics 123(1): 286–293.

Fisher, M. L. 1985. An applications oriented guide to lagrangian
relaxation. Interfaces 15(2): 10–21.
Fonarow, G. C., L. W. Stevenson, J. A. Walden, N. A. Livingston,
A. E. Steimle, M. A. Hamilton, J. Moriguchi, J. H. Tillisch,
M.A. Woo. 1997. Impact of a comprehensive heart failure
management program on hospital readmission and functional
status of patients with advanced heart failure. J. Am. Coll. Car-
diol. 30(3): 725–732.
Foster, D., G. Harkness. 2010. Healthcare Reform: Pending Changes
to Reimbursement for 30-day Readmissions. Thomson Reuters,
Ann Arbor, MI.
Geoffrion, A. M. 1974. Lagrangean Relaxation for Integer Program-
ming. Springer, Berlin.
Gilks, W. R., P. Wild. 1992. Adaptive rejection sampling for gibbs
sampling. Appl. Stat. 41(2): 337–348.
Gocgun, Y., A. Ghate. 2012. Lagrangian relaxation and constraint
generation for allocation and advanced scheduling. Comput.
Oper. Res. 39(10): 2323–2336.
Gonseth, J., P. Guallar-Castillon, J. R. Banegas, F. Rodrıguez-Arta-
lejo. 2004. The effectiveness of disease management pro-
grammes in reducing hospital re-admission in older patients
with heart failure: A systematic review and meta-analysis of
published reports. Eur. Heart J. 25(18): 1570–1595.
Gordon, S. C. 2002. Stochastic dependence in competing risks.
Am. J. Polit. Sci. 46: 200–217.
Graham, J., J. Tomcavage, D. Salek, J. Sciandra, D. E. Davis, W. F.
Stewart. 2012. Postdischarge monitoring using interactive
voice response system reduces 30-day readmission rates in a
case-managed medicare population. Med. Care 50(1): 50.
Gustafson, P. 1997. Large hierarchical bayesian analysis of multi-
variate survival data. Biometrics 53: 230–242.
Halfon, P., Y. Eggli, I. Prêtre-Rohrbach, D. Meylan, A. Marazzi, B.
Burnand. 2006. Validation of the potentially avoidable hospi-
tal readmission rate as a routine indicator of the quality of
hospital care. Med. Care 44(11): 972–981.
Held, M., P. Wolfe, H.P. Crowder. 1974. Validation of subgradient
optimization. Math. Program. 62(6): 1265–1282.
Helm, J. E., M. P. Van Oyen. 2014. Design and optimization meth-
ods for elective hospital admissions. Oper. Res. 6(1): 62–88
Helm, J. E., M. P. Van Oyen, T. R. Rohleder. 2013. Priority sched-
uling in a queueing network with an application to itinerary
completion at destination medical centers. Working paper,
Mayo Clinic.
Hernandez, A. F., M. A. Greiner, G. C. Fonarow, B. G. Hammill,
P. A. Heidenreich, C. W. Yancy, E. D. Peterson, L. H. Curtis.
2010. Relationship between early physician follow-up and 30-
day readmission among medicare beneficiaries hospitalized
for heart failure. JAMA 303(17): 1716–1722.
Hougaard, P., P. Hougaard. 2000. Analysis of Multivariate Survival
Data, Vol. 564. Springer New York.
Hu, M., B. L. Jacobs, J. S. Montgomery, C. He, J. Ye, Y. Zhang, J.
Brathwaite, T. M. Morgan, K. S. Hafez, A. Z. Weizer, M. Hu,
B. L. Jacobs, J. S. Montgomery, C. He, J. Ye, Y. Zhang,
J. Brathwaite, T. M. Morgan, K. S. Hafez, A. Z. Weizer,
S. M. Gilbert, C. T Lee, M. S. Lavieri, J. E. Helm,
B. K. Hollenbeck, T. A. Skolarus. 2014. Sharpening the focus
on causes and timing of readmission after radical cystectomy
for bladder cancer. Cancer 120(9): 1409–1416.
Ibrahim, J. G., M. Chen, D. Sinha. 2005. Bayesian Survival Analysis.
Wiley Online Library, Hoboken, NJ.
Jack, B. W., V. K. Chetty, D. Anthony, J. L. Greenwald, G. M. San-
chez, A. E. Johnson, S. R. Forsythe, J. K. O’Donnell, M. K.
Paasche-Orlow, C. Manasseh, et al. 2009. A reengineered hos-
pital discharge program to decrease rehospitalization: A ran-
domized trial. Ann. Intern. Med. 150(3): 178.
Kansagara, D., H. Englander, A. Salanitro, D. Kagen, C. Theobald,
M. Freeman, S. Kripalani. 2011. Risk prediction models for
hospital readmission: A systematic review. JAMA 306(15):
1688–1698.
Keller, J. B. 1974. Optimum checking schedules for systems sub-
ject to random failure. Manage. Sci. 21(3): 256–260.
Luss, H. 1976. Maintenance policies when deterioration can be
observed by inspections. Oper. Res. 24(2): 359–366.
McHugh, M. D., C. Ma. 2013. Hospital nursing and 30-day read-
missions among medicare patients with heart failure, acute
myocardial infarction, and pneumonia. Med. Care 51(1): 52–59.
Melton, L. D., C. Foreman, E. Scott, M. McGinnis, M. Cousins.
2012. Prioritized post-discharge telephonic outreach reduces
hospital readmissions for select high-risk patients. Am. J.
Manag. Care 18(12): 838.
Minott, J. 2008. Reducing hospital readmissions. Acad. Health 23
(2): 1–10.
Oakes, D. 1989. Bivariate survival models induced by frailties. J.
Am. Statist. Assoc. 84(406): 487–493.
Pan, S. J., Q. Yang. 2010. A survey on transfer learning. IEEE
Trans. Knowl. Data Eng. 22(10): 1345–1359.
Pang-Ning, T., M. Steinbach, V. Kumar, et al. 2006. Introduction
to data mining. Library of Congress.
Richards, F. J. 1959. A exible growth function for empirical use. J.
Exp. Bot. 10(2): 290–301.
Rosen, A. K., S. Loveland, M. Shin, M. Shwartz, A. Hanchate, Q.
Chen, H. M. A. Kaafarani, A. Borzecki. 2013. Examining the
impact of the AHRQ patient safety indicators (psis) on the
veterans health administration: The case of readmissions.
Med. Care 51(1): 37–44.
Rosenberg, M. A., E. W. Frees, J. Sun, P. H. Johnson, J. M. Robin-
son. 2007. Predictive modeling with longitudinal data: A case
study of Wisconsin nursing homes. N Am. Actuar. J. 11(3): 54.
Sokal, R. R. 1958. A statistical method for evaluating systematic
relationships. Univ. Kans. Sci. Bull. 38: 1409–1438.
Spiegelhalter, D., A. Thomas, N. Best, D. Lunn. 2003. WinBUGS
user manual version 1.4 (January 2003). Available at http://
www.politicalbubbles.org/bayes_beach/manual14.pdf
de Toledo, P., S. Jimenez, F. del Pozo, J. Roca, A. Alonso, C. Her-
nandez. 2006. Telemedicine experience for chronic care in
copd. IEEE Trans. Inf. Technol. Biomed. 10(3): 567–573.
Wallmann, R., J. Llorca, I. Gomez-Acebo, A. C. Ortega, F. R. Rol-
dan, T. Dierssen-Sotos. 2013. Prediction of 30-day cardiac-
related-emergency-readmissions using simple administrative
hospital data. Int. J. Cardiol. 164(2): 193–200.
van Walraven, C., I. A. Dhalla, C. Bell, E. Etchells, I. G. Stiell, K.
Zarnke, P. C. Austin, A. J. Forster. 2010. Derivation and vali-
dation of an index to predict early death or unplanned read-
mission after discharge from hospital to the community. Can.
Med. Assoc. J. 182(6): 551–557.
Watson, A. J., J. O’Rourke, K. Jethwani, A. Cami, T. A. Stern, J. C.
Kvedar, H. C. Chueh, A. H. Zai. 2011. Linking electronic
health record-extracted psychosocial data in real-time to risk
of readmission for heart failure. Psychosomatics 52(4):
319–327.
Wolinsky, F., S. Bentler, E. Cook, E. Chrischilles, L. Liu, K.
Wright, J. Geweke, M. Obrizan, C. Pavlik, R. Ohsfeldt, et al.
2009. A 12-year prospective study of stroke risk in older
medicare beneficiaries. BMC Geriatr. 9(1): 17.
Yu, S., A. van Esbroeck, F. Farooq, G. Fung, V. Anand, B. Krish-
napuram. 2013. Predicting readmission risk with institution
specific prediction models. 2013 IEEE International Confer-
ence on Healthcare Informatics (ICHI), IEEE, Philadelphia,
PA, pp. 415–420.

Prognostics with optimization

More Related Content

What's hot (15)

Similar to Prognostics with optimization (20)

More from John Frias Morales, DrBA, MS (12)

Recently uploaded (20)

Prognostics with optimization