SFSCON23 - Gianluigi Greco - AI Algorithms for Digital Therapeutics

AI Algorithms for Digital Therapeutics
Prof. Gianluigi Greco – Head of the Department of Mathematics and Computer Science, University of Calabria
President of the Italian Association for Artificial Intelligence

The role of artificial intelligence
in the digital medicine

THE DIGITAL TRANSFORMATION
Digital Health
Technologies, platforms and systems that engage users for purposes related
to lifestyle improvement, well-being and health. Other objectives may be to
acquire, store or transmit health data, or to support clinical activities. To be
deployed, applications in this area do not require clinical trials, nor
regulatory supervision from any type of national or international body

Digital Health
Digital Medicine
Software and hardware for clinical measurements and/or to intervene
directly on health. They require clinical efficacy trials and are typically
classified as medical devices

THEORY AND PRACTICE
The digital medina is based on software and
hardware solutions, which do not necessarily
have to use artificial intelligence techniques
THEORY

THEORY AND PRACTICE
The digital medina is based on software and
hardware solutions, which do not necessarily
have to use artificial intelligence techniques
THEORY
+AI
512
DEVICES
cardiovascular
radiology
THEORY
PRACTICE

ALGORITHMS IN MEDICINE
The number of algorithms using Artificial Intelligence in
the medical field has increased 10-fold in the last year!
x10
More and more applications

ALGORITHMS IN MEDICINE
The number of algorithms using Artificial Intelligence in
the medical field has increased 10-fold in the last year!
x10
More and more applications
superuman
performances

178 miliardi $
8 miliardi $
AI&HEALTH MARKET

Surveillance: AI can help identify specific demographics or
geographical locations where the prevalence of disease or
high-risk behaviours exist; it can also help to implement digital
epidemiological surveillance
Public health
Global health: AI may provide opportunities to address health
challenges in low-and middle-income countries (LMICs); these
challenges include acute health workforce shortages and weak
public health surveillance systems;
DOMAINS FOR AI IN HEALTCARE
Drug discovery: recently developed AI approaches provide new
solutions to enhance the efficacy and safety evaluation of
candidate drugs based on big data modelling and analysis;

Clinical research: mainstream medical knowledge resources
are already using ML algorithms to rank search results, including
algorithms that learn from users’search behaviour;
Biomedical reserach
Personalized medicine: strongly relies on a scientific understanding
of how an individual patient's unique characteristics, such as
molecular and genetic profiles, make this patient vulnerable to a
disease and sensitive to a therapeutic treatment;

Healthcare systems are characterized by a heavy administrative
workflow with a wide range of actors and institutions,
comprising patients (e.g. management of billing), health
professionals, healthcare facilities and organisations (e.g.
patient flow), imaging facilities, laboratories (e.g. supply chain
of consumables), pharmacies, payers, and regulators.
AI can perform routine and burocratic tasks in a more
efficient, accurate and unbiased fashion
Applications to scheduling, patient flow management,
identification of fraudulent activities
Health administration

Radiology and digital pathology: segmentation with
limited human supervision to automatically localise and
delineate the boundaries of anatomical structures or lesions;
Clinical Practice
Emergency medicine: improve patient prioritisation during
triage; organisational planning and management within the
emergency department;
Surgery: integration of diverse sources of information (patient
risk factors, anatomic information, etc.) in the development of
better surgical decisions;
Home care: self-management of chronic diseases and diseases
that affect the elderly;

AI&Health @Unical
Carlo Adornetto Pierangela Bruno Gianluigi Greco Giuseppe Covello Vincenzo Rizzuto

In fields such as Engineering, Chemistry and Physics, the design of device
structures is progressively supported by Deep Learning methods
Objective: design materials, devices or tools based on the properties they
should exhibit
Conventional Approach
Inverse Design
Properties
INVERSE DESIGN

PROPERTIES OF INTEREST
Functional Properties:
▪ Bind to a specific target receptor or enzyme
Physicochemical Properties:
▪ Molecule's chemical structure and behavior, such as solubility, stability,
boiling point, melting point, and chemical reactivity
Toxicity and Safety:
▪ In a biological or environmental context, it's important to consider
molecules toxicity and safety profile
Specific Target or Application-Related Properties:
▪ For instance, if you're designing a molecule for use in a particular type of
semiconductor, you would need to focus on properties relevant to that
application, like charge carrier mobility or bandgap

ISSUES
× Non-uniqueness of the solution

ISSUES
× Drastically different molecules can
produce very similar responses.
Design Space Properties Space

ISSUES
× High dimensionality of the design space

ISSUES
× High dimensionality of the design space
× Feasibility constraints on the design

FEASIBILITY CONSTRAINTS
Chemical Feasibility:
Adhering to valence rules, satisfying octet rules for most atoms, and avoiding strained
or unstable configurations.
Synthetic Accessibility:
The designed molecule should be synthesizable using available or reasonable
synthetic methods. Complex or exotic reactions and reagents may be impractical,
costly, or impossible to implement.
Reaction Conditions:
Extremely high temperatures, pressures, or toxic reagents can be prohibitive.
Hazardous or toxic materials and reactions should be avoided or properly managed.
Short shelf life or chemical instability can be problematic.
Cost:
High production costs can make a molecule economically unviable.
Regulatory and Compliance Constraints:
Compliance with safety, environmental, and legal regulations is essential.
Purity and Characterization:
The designed molecule should be synthesizable with a high degree of purity, and
methods for characterizing and quality control should be established.

STATE OF THE ART
Feed Forward Tandem
cGAN
Output-DEPENDENT
Output-INDEPENDENT
NOT requiring fine-tuning on the desired Requiring fine-tuning on the desired
Direct Inverse Forward Simulator
VAE-Based
Fine tuning on
Adjoint
Train on Dataset of pairs
cVAE
GLOnets
Discriminator
Encoder/Decoder

STATE OF THE ART
Feed Forward Tandem
cGAN
Output-DEPENDENT
Output-INDEPENDENT
NOT requiring fine-tuning on the desired Requiring fine-tuning on the desired
Direct Inverse Forward Simulator
VAE-Based
Fine tuning on
Adjoint
Train on Dataset of pairs
cVAE
GLOnets
Discriminator
Encoder/Decoder
Most architectures work at the level of the original (highly
dimensional) design space
No feasibility constraint considered in the design process
Random init to start the exploration/optimization in the
design space

GIDnet – IJCAI 2023
We embed the design space into a suitably-defined latent space to deal with complex
representations going beyond plain numerical values

We restrict the latent design space to feasible regions
Reconstruction Loss To enforce constraints
Softmax

Rather than using a «blind» generator that tries to compute the solution by starting
from some random initialization, we start the exploration of the latent space by first
looking at the dataset and identifying educated guesses called seeds:

We provide an exploration mechanism togheder with a mechanism (Selection Layer) such that the network can
automatically choose a starting point for the exploration, as one of the given seeds, or alternatively, as a linear
combination of the seeds
Push the selection layer towards a
single choice
Regularization term for constraints
seeds

graphene
5 layers thin film metamaterials
▪ each layer with thickness within the range 1-60 nm
▪ material can be Ag, Al2O3, ITO, Ni, or TiO2
▪ we have to represent its thickness plus the material as a one-
hot encoding over 5 alternatives.
▪ Each structure is associated with reflectance and transmittance
spectra, obtained via the transfer matrix method simulated on
an infinite glass substrate, for two polarizations, at the incident
angles of 25, 45, and 65 degrees, for 200 equally spaced points
over the range 450-950 nm
Transfer Matrix
Method

▪ In disease, cells genes are often under-expressed
or over-expressed.
▪ High-throughput sequencing and Microarrays
are efficient techniques to gather data that can
be used to determine the expression pattern of
thousands of genes
SETTING

▪ In disease, cells genes are often under-expressed
or over-expressed.
▪ High-throughput sequencing and Microarrays
are efficient techniques to gather data that can
be used to determine the expression pattern of
thousands of genes
SETTING
Course of Dimensionality
Thousands of genes for few patients (Linear dep. between genes)
Noise and redundancy
Data collection is often multi-centric and carried out with
heterogeneous devices
Class Imbalance
Sequencing mostly takes place on pathological patients
patients
genes

SETTING
▪ Chronic Lymphocytic Leukemia (CLL) is a hematologic neoplasm characterized by
an accumulation of lymphocytes in the blood, bone marrow, and lymphatic
organs (lymph nodes and spleen)
▪ In more than half of the patients, CLL is diagnosed incidentally, and some
patients can remain stable for more than 10 years, while others may experience
rapid worsening
▪ Currently, it is not possible to establish precise rules for the prevention of CLL, as
its causes are not completely clear
DIAGNOSIS EVENT
Worsening of the disease:
Death or Need of Trearment
TTFT
Time to first treatment

FEATURE SELECTION
• Main Ideas:
• Iterative Procedure which selects a set
of meaningful genes at each iteration
• Neural approach to filter redundant
genes into the genes space
• Ad-hoc defined EXPLAINABLE AI-
based method to select the most
impactful genes
Correlation Clustering
Neural Filtering
Neural Network for
prediction
SHAP XAI Selection
The process of selecting a subset of relevant features (variables, predictors) to use in
predictive models, to reduce the computational cost and to improve the performance

A CLOSER LOOK AT THE STEPS
Neural Filtering
Neural Network for prediction
SHAP XAI Selection
(of genes)

Neural Filtering
SHAP XAI Selection
Clusters k1 . . . kq

Neural Filtering
SHAP XAI Selection
Clusters k1 ... kq
q Genes

Neural Filtering
SHAP XAI Selection
Select and Save the most meaningful genes according an
ad-hoc defined
SHAP-based score
Filter the genes according to the correlation between
SHAP values and genes values
Clusters k1 ... kq
q Genes

APPLICATION TO CLL
Genes Selection using Deep Learning and Explainable Artificial Intelligence for Chronic
Lymphocytic Leukemia Predicting the Need and Time to Therapy
Fortunato Morabito, Carlo Adornetto, Paola Monti, Adriana Amaro, Francesco
Reggiani, Monica Colombo, Yissel Rodriguez-Aldana, Giovanni Tripepi, Graziella
D'Arrigo, Claudia Vener, Federica Torricelli, Rossi Teresa, Manlio Ferrarini, Giovanna
Cutrona, Antonino Neri, Massimo Gentile and Greco Gianluigi
O-CLL Dataset
TTFT
EVENT
19.367 GENES
Pz1
Pz 217
Yes
Yes
No
No
< 24
> 24
…
…
…
…
…
97 patients
months

APPLICATION TO CLL
EVENT
GENE XAI score % Corr.
SL4A1 11,79 -0,98
CAES 9,78 -0,97
VS20 6,91 -0,97
CT1 6,16 0,97
HTD4 5,94 0,96
FADD 3,83 0,95
GNE 3,35 0,97
PIGP 3,25 0,96
Eight of the top ten genes selected by the algorithm were found in the Reactome
pathway database, showing an involvement in various pathways such as signal
transduction, gene expression (transcription), protein metabolism, immune
system, cell cycle and apoptosis.
7 of them are involved in protein-protein interaction (PPI)
Iteration C.I. 95%
1 77.2 – 92.7
2 74.7 – 91.2
3 68.6 – 89.3
4 64.3 – 83.6
Final 79.1 – 92.9
TTFT

How to quantify the effect of blepharoplasty on rejuvenation?
OPHTALMOPLASTIC SURGERY

In the evaluation and management of ptosis (by blepharoplasty), measurements of MRD1,
MRD2, and LF are time-consuming, subjective, and prone to human error.
Correlation up to

AI to quantify proptosis and identify patients to be treated surgically
Case study of 56 paediatric patients (31 of whom were surgically treated)

Generative AI to predict the aesthetic outcome of surgery
109 Pazienti
Pre-op Post-op
Dataset of patients undergoing Orbital
Decompression for Thyroid-Associated
Ophthalmopathy (TAO)

• There are no tools generated to predict Post-op
• Requires a precision, patient-tailored approach
• Lots of individual variables to consider
• AI-based rendering to generate post-operative
images from pre-operative photo
• Explore the results for different surgical variables
The Problem
The Idea
How?
Pre-op Post-op
Tipo Intervento
Tipo Sututa
Pre-op
Post-op
Post-op
Pre-op
1. Data Collection (smartphone + clinical data)
Post-op
2. Generative .AI
OPHTALMOPLASTIC SURGERY @Unical

Utilities and Advantages:
To allow ophthalmoplastic surgeons to have patients who have to undergo surgery see a prediction of
surgical outcomes through a rendering generated on the basis of the data provided. Each rendering will
have the ability to be customized by changing the surgical variables.
Features & Functionality:
- Mobile platform
- Ability to take images with AI support to standardize photo
- Possibility to customize the intervention sheet for each individual patient
- Obtain more renderings of the surgery according to the surgical plan
- Iniziali:
-Età:
-Etnia
-Anno di nascita:
-Quadro clinico:
-Tipologia di ptosi:
-Causa della ptosi:
Dati relativi all’intervento:
-Tipologia di intervento:
-Tipologia di sutura:
OPHTALMOPLASTIC SURGERY @Unical

Digital Health
Digital Medicine
Software and hardware for clinical measurements and/or to intervene
directly on health. They require clinical efficacy trials and are typically
classified as medical devices
DTx
Software that delivers therapeutic interventions to prevent, manage or treat a
medical disorder or disease. Clinical evidence and Real World Evidence are required

Lenire [FDA 03/2023]
Treating Tinnitus Symptoms
It uses the principle of bimodal neuromodulation: it
provides mild electrical impulses to the tongue
combined with sound reproduced through
headphones to drive long-term changes or
neuroplasticity in the brain to treat tinnitus.
Clinical trial on 112 patients and RWE of 204 users,
with 79.4% improvement
The frequencies of the sounds played by the
headphones must be customized

Tinnitracks [non-FDA]
Treating Tinnitus Symptoms
Filter music based on the specific tinnitus frequency
Clinical trial on 98 patients, with 65% improvement
The application filters the music played in real time,
optimizing it according to the frequency of tinnitus

CognICA [FDA 10/2021]
Assessment of cognitive functions
Implement a rapid test, based on displaying images
at a rapid pace on the iPad screen and asking you
to identify them as animals or non-animals
Clinical trial on 91 patients, with 94% accuracy
The classification of the patient, based on
response speed and accuracy, takes place with
a regression system

CanvasDx [FDA 06/2021]
Early diagnosis of autism spectrum disorders
An app collects behavioral data, videos and
feedaback from healthcare professionals
Clinical trial on 425 children
The application classifies the various cases
through a machine learning algorithm
(gradient boosted decision tree algorithm)

MedRhythm [FDA Breakthrough Device Designation 2020]
Walking in stroke patients
Use of rhythmic auditory stimulation, to facilitate
walking and the ability to synchronize movements
Clinical trial on 11 patients
The application analyzes the cadence and
quality of walking and dynamically adapts
auditory stimuli

Pivot Breathe [FDA 10/2017]
Tobacco addiction
Coaching app that uses a sensor to read carbon
monoxide levels in the breath
Clinical trial on 319 patients, 35% success
The virtual coaching system is adaptive and
personalizes the experience based on the
user's characteristics

RhythmAnalytics [FDA 5/2019]
Identification of cardiac arthymia
System that allows you to detect and analyze
cardiographic traces, and other data collected with sensors
Testing on 120,000 episodes, with superior reliability
to expert panels
The artimia identification system is based on
anomaly identification mechanisms, trained
on millions of cases

Insulia [FDA 7/2021]
Treatment of type II diabetes
Integrated system of sensors and personalized
recommendation, about the correct dosage of insulin
Clinical trial involving 191 patients
Expert system that encodes domain knowledge,
between recorded values and recommended doses

RelieVRx [FDA 7/2021]
Treatment for chronic pain reduction
It includes a VR headset and a device that amplifies
the sound of the user's breathing to assist in
breathing exercises. Use the principles of cognitive
behavioral therapy (digital CBT)
Clinical trial on 188 patients, efficacy on 65%
Use of immersive environment, with strong
characteristics of human-machine interaction

EndeavorRx [FDA 6/2020]
Improved attention in children 8-12 years
Video game that develops specific cognitive areas,
which require particular stress in children with
attention deficit hyperactivity disorder (ADHD)
Clinical trial on 600 children, success on 73%
Videogames are among the most classic development
environments for artificial intelligence technologies

AI IN DIGITAL DEVICES
DATABASE
BACK-END
Data and logs

DATABASE
BACK-END
analytics
corrective therapy
APP
Data and logs

DATABASE
BACK-END
analytics
corrective therapy
Data and logs
Lifestyle and sensor data
APP
personalized reccomandations
APP

SENSORS
DATABASE
BACK-END
analytics
corrective therapy
Data and logs
APP
APP

AI
SENSORS
DATABASE
BACK-END
analytics
corrective therapy
Data and logs
APP
APP

AI regulations in digital medicine

REGULATIONS
Artificial Intelligence
for
Medical Devices

REGULATIONS
[..] AI system shall be considered high-risk where both
of the following conditions are fulfilled:
▪ the AI system is intended to be used as a safety
component of a product, or is itself a product,
covered by the Union harmonisation legislation
listed in Annex II;
▪ the product whose safety component is the AI
system, or the AI system itself as a product, is
required to undergo a third-party conformity
assessment with a view to the placing on the
market or putting into service of that product
pursuant to the Union harmonisation legislation
listed in Annex II.
In addition to the high-risk AI systems referred to in
paragraph 1, AI systems referred to in Annex III shall
also be considered high-risk

LEGISLATION ON MEDICAL DEVICES
Regulation (EU) 2017/745 of the European Parliament and
of the Council of 5 April 2017 on medical devices MDR
Regulation (EU) 2017/746 of the European Parliament and of the
Council of 5 April 2017 on in vitro diagnostic medical devices IVDR
The Medical Devices Regulation applies since 26 May 2021
The In Vitro Diagnostic Devices Regulation applies since 26 May 2022

MDR GENERAL ARCHITECTURE
- In the European Union (EU) MEDICAL DEVICES must undergo a conformity assessment to
demonstrate they meet legal requirements to ensure they are safe and perform as intended.
- They are regulated at EU Member State level, but the European Medicines Agency (EMA) is
involved in the regulatory process.
- Manufacturers can place a CE (Conformité Européenne) mark on a medical device once it
has passed a conformity assessment.
- The conformity assessment usually involves an audit of the manufacturer's quality system and,
depending on the type of device, a review of technical documentation from the
manufacturer on the safety and performance of the device.
- EU Member States designate accredited notified bodies to conduct conformity
assessments. For certain high-risk devices, notified bodies shall request the opinion of specific
expert panels before issuing the certificate of conformity. These expert panels benefit from
EMA's technical and scientific support.

MDR AND AI
Classification rule 11
Software intended to provide information which is used to take decisions with
diagnosis or therapeutic purposes is classified as class IIa, except if such decisions
have an impact that may cause:
▪ Death or an irreversible deterioration of a person's state of health, in which case it
is in class III; or
▪ Serious deterioration of a person's state of health or a surgical intervention, in
which case it is classified as class IIb.
Software intended to monitor physiological processes is classified as class IIa,
except if it is intended for monitoring of vital physiological parameters, where the
nature of variations of those parameters is such that it could result in immediate
danger to the patient, in which case it is classified as class IIb.
All other software are classified as class I.

MDR AND AI
▪ Dificulties in defining appropriate
classification for medical devices

MDR AND AI
▪ Dificulties in defining appropriate
classification for medical devices
▪ AI systems are often black box and implement various functionalities
▪ Guidelines for softwares made of different modules
▪ Examples and pratical guidance
MDR classes are related to risk categories

FURTHER ISSUES WITH DTx
- Although in some European countries several procedures
have been set up for the marketing authorization and
reimbursement of DTx, these are just individual and
uncoordinated initiatives.
- The European regulatory system concerning DTx is still
immature and specific regulations aimed at evaluating
these tools and ensuring the safety of the devices and the
integrity of the data collected are lacking.
- It is necessary to define specific indications for DTx
approval, taking into account their peculiar characteristics,
such as the rapidity of digital evolution and the issues
concerning patients’ privacy and data security

A FEW EXCEPETIONS
- A specific fast-track regulatory process
for digital health applications (DiGA,
Digitale Gesundheitsanwendungen)
was launched
- Physicians can prescribe digital health
applications (including several DTx)
that can then be reimbursed by health
insurance companies
- The National Institute for Health and Care
Excellence (NICE) established an Office for
Digital Health to accelerate efforts to deliver
innovation to the health and care system
- The Office launched a project, the
Innovative Devices Access Pathway (IDAP),
to design and scope an innovative access
pathway for selected medical and digital
health technologies

SOLUTION SECURITY
As with any device that handles patient data, privacy and
security are crucial, especially those with frequent software
updates or connections to other electronic health record
devices or systems.

BALANCING DATASETS
It is very complex to evaluate the effectiveness of machine
learning algorithms in different patient populations. Training
datasets must be well defined, and address bias issue.
devices or systems.

With current platforms for prototyping AI solutions, anyone can
develop a digital app and make it available on a commercial
platform.
COST-EFFECTIVE TECHNIQUES AND TOOLS
It is very complex to evaluate the effectiveness of machine
learning algorithms in different patient populations. Training
datasets must be well defined, and address bias issue.
devices or systems.

With current platforms for prototyping AI solutions, anyone can
develop a digital app and make it available on a commercial
platform.
COST-EFFECTIVE TECHNIQUES AND TOOLS

• Noise and artefacts in AI’s input or data shifts between training and real-world
can cause false diagnosis and/or incorrect scheduling or prioritization
Keep human in the loop, and implement approaches
that continuously improve over time
Patient harm due to AI errors
FURTHER RISKS OF AI

• Noise and artefacts in AI’s input or data shifts between training and real-world
can cause false diagnosis and/or incorrect scheduling or prioritization
Patient harm due to AI errors
Keep human in the loop, and implement approaches
that continuously improve over time
Misuse of medical AI tools
▪ Lack of AI training of medical staff can cause incorrect usage of tools, resulting in
incorrect medical assessment and decision making (again with potential harm)
Invest on education, training and res-skilling
programs, targeted to digital medicine
FURTHER RISKS OF AI

Thank You!
Visit our web site
https://aixia.it/en/
Join us
https://aixia.it/en/diventa-socio/
Follow us
LinkedIn – https://it.linkedin.com/company/aixia1988
Twitter – @AI_x_IA
Facebook – @AIxIA1988

SFSCON23 - Gianluigi Greco - AI Algorithms for Digital Therapeutics

More Related Content

Similar to SFSCON23 - Gianluigi Greco - AI Algorithms for Digital Therapeutics (20)

More from South Tyrol Free Software Conference (20)

Recently uploaded (20)

SFSCON23 - Gianluigi Greco - AI Algorithms for Digital Therapeutics