Schedule

(CET)
Monday
10/01/2022
9:00
-
9:30
Organizers' speech

Organizers' speech

10/01/2022 - 9:00 - 9:30
9:30
-
11:00
Stanley Durrleman
Part 1. Modeling and predicting disease progression with longitudinal data

Stanley Durrleman
Part 1. Modeling and predicting disease progression with longitudinal data

10/01/2022 - 9:30 - 11:00

In this course, we will present new methods for learning disease progression model from longitudinal data sets. Longitudinal data are ubiquitous in medicine: they contain observations of patients followed at multiple time-points.

These data sets raise several challenges: the number of patients is usually small, the spacing between time-points is irregular, observation period cover only a part of the disease stages. Moreover, for some diseases like neurodegenerative diseases, age / time is not a good marker of disease progression, which further adds to the difficulty.

We will start by reviewing the classical statistical approach for this type of problems, where mixed effect models for repeated measurements play a central role. We will then discuss the pros and cons of such methods, and justify the need for the development of new methods. We will introduce a new technique called disease course mapping which learns distributions of disease trajectories, where trajectories are seen as curves on the data space. The method lies at the interface of Riemanian geometry, dynamical systems, and statistical learning.

We will show how these techniques can be used to forecast disease progression in new patients, and design more powered clinical trials. We will use neurodegenerative diseases as key application area with examples taken in the fields of Alzheimer disease, Parkinson disease and Huntington disease.

Basics concepts in statistical inference and mixed effect modelling are a pre-requisite. Knowledge in geometry will help, although the main concepts will be introduced during the lecture.

11:00
-
11:30
Break

Break

10/01/2022 - 11:00 - 11:30
11:30
-
12:30
Stanley Durrleman
Part 2. Modeling and predicting disease progression with longitudinal data

Stanley Durrleman
Part 2. Modeling and predicting disease progression with longitudinal data

10/01/2022 - 11:30 - 12:30

In this course, we will present new methods for learning disease progression model from longitudinal data sets. Longitudinal data are ubiquitous in medicine: they contain observations of patients followed at multiple time-points.

These data sets raise several challenges: the number of patients is usually small, the spacing between time-points is irregular, observation period cover only a part of the disease stages. Moreover, for some diseases like neurodegenerative diseases, age / time is not a good marker of disease progression, which further adds to the difficulty.

We will start by reviewing the classical statistical approach for this type of problems, where mixed effect models for repeated measurements play a central role. We will then discuss the pros and cons of such methods, and justify the need for the development of new methods. We will introduce a new technique called disease course mapping which learns distributions of disease trajectories, where trajectories are seen as curves on the data space. The method lies at the interface of Riemanian geometry, dynamical systems, and statistical learning.

We will show how these techniques can be used to forecast disease progression in new patients, and design more powered clinical trials. We will use neurodegenerative diseases as key application area with examples taken in the fields of Alzheimer disease, Parkinson disease and Huntington disease.

Basics concepts in statistical inference and mixed effect modelling are a pre-requisite. Knowledge in geometry will help, although the main concepts will be introduced during the lecture.

12:30
-
14:00
Lunch Break

Lunch Break

10/01/2022 - 12:30 - 14:00
14:00
-
14:30
Organizers' speech

Organizers' speech

10/01/2022 - 14:00 - 14:30
14:30
-
15:30
I.Glenn Cohen - FREE ACCESS
AI/ML in medicine: Legal and Ethical Issues

I.Glenn Cohen - FREE ACCESS
AI/ML in medicine: Legal and Ethical Issues

10/01/2022 - 14:30 - 15:30

This session will cover the legal and ethical issues that arise in building and applying AI/ML. It will consider the stages from data acquisition to ultimate deployment in clinical settings. Issues covered may include privacy, trade secrecy, liability, informed consent, bias and discrimination, adaptive and locked algorithms, etc.

15:30
-
16:30
Ran Balicer - FREE ACCESS
Using health data and AI to drive healthcare transformation in practice

Ran Balicer - FREE ACCESS
Using health data and AI to drive healthcare transformation in practice

10/01/2022 - 15:30 - 16:30

TBD

16:30
-
17:00
Break

Break

10/01/2022 - 16:30 - 17:00
17:00
-
18:30
Yoshua Bengio - FREE ACCESS
AI for Drug Discovery

Yoshua Bengio - FREE ACCESS
AI for Drug Discovery

10/01/2022 - 17:00 - 18:30

Researchers are investigating the use of machine learning in the whole pipeline of drug discovery, from the basic research on cell biology to discover candidate targets to the exploration of the space of drugs and to increased efficiency of clinical trials. Thanks to progress in biotechnology and the establishment of public databases regarding molecular biology, the amount of data that is collected about what is going on inside our cells is rapidly increasing, and for example synthetic biology makes it possible to synthesize and assay huge numbers of candidate proteins in a single experiment. It is likely that these advances will enable a major transformation in molecular biology, cell biology and drug discovery thanks to machine learning. The most remarkable recent success is that of modeling the 3-dimensional structure of proteins. Looking forward, machine learning can be used to help us model very complex biological or chemical phenomena for which we do not yet have full mechanistic understanding. Using these models, which can be causally structured, machine learning can also help us search the space of interventions (such as drugs) which can achieve desirable therapeutic results. The kind of machine learning methods necessary for such an exploration is different from the traditional supervised learning with a fixed dataset. It is a form of active learning where the learner, like a scientist, can propose experiments to be performed in order to acquire information and improve the model so that better interventions can be discovered downstream. We will present new machine learning ideas incorporating reinforcement learning and generative active learning aimed at this particular setting. 

Tuesday
11/01/2022
9:00
-
10:30
Wendy W.Chapman
Part 1. Looking under the hood of the Ferrari: how can we extract information from clinical notes?

Wendy W.Chapman
Part 1. Looking under the hood of the Ferrari: how can we extract information from clinical notes?

11/01/2022 - 9:00 - 10:30

Most of the data collected about patients during their healthcare visits are text. That means that most patients’ symptoms, social risk factors, medical history, and family history are inaccessible to AI applications for decision support, predictive analytics, surveillance, and research. Deep learning has shown impressive performance in machine translation and other NLP application areas, and is beginning to demonstrate success in the clinical space. However, there is still a need for rules-based and traditional machine learning approaches. We will provide examples of how NLP is being used to support health, step through a case study to learn first-hand what makes NLP challenging in this domain, and demonstrate how a rule-based system can be used to develop NLP tools potentially useful for clinical text. 

10:30
-
11:00
Break

Break

11/01/2022 - 10:30 - 11:00
11:00
-
12:30
Wendy W.Chapman
Part 2. Looking under the hood of the Ferrari: how can we extract information from clinical notes?

Wendy W.Chapman
Part 2. Looking under the hood of the Ferrari: how can we extract information from clinical notes?

11/01/2022 - 11:00 - 12:30

Most of the data collected about patients during their healthcare visits are text. That means that most patients’ symptoms, social risk factors, medical history, and family history are inaccessible to AI applications for decision support, predictive analytics, surveillance, and research. Deep learning has shown impressive performance in machine translation and other NLP application areas, and is beginning to demonstrate success in the clinical space. However, there is still a need for rules-based and traditional machine learning approaches. We will provide examples of how NLP is being used to support health, step through a case study to learn first-hand what makes NLP challenging in this domain, and demonstrate how a rule-based system can be used to develop NLP tools potentially useful for clinical text. 

12:30
-
14:00
Lunch break

Lunch break

11/01/2022 - 12:30 - 14:00
14:00
-
15:00
Use case - Stéphanie Allassonnière & Julien Stirneman
A decision support system for prenatal diagnosis of fetal anomalies by ultrasound using statistical learning

Use case - Stéphanie Allassonnière & Julien Stirneman
A decision support system for prenatal diagnosis of fetal anomalies by ultrasound using statistical learning

11/01/2022 - 14:00 - 15:00

In this work, we propose a method to build a decision support system for the diagnosis of rare diseases in the fetus. We aim to minimise the number of medical tests required to reach a state where the uncertainty about the patient's disease is below a predetermined threshold. To solve this optimization task, we propose an operational reinforcement learning algorithm for our very high-dimensional problem where the learned strategies perform much better than classical greedy strategies. We also present a way to combine expert knowledge, expressed as conditional probabilities, with clinical data. This is crucial because the scarcity of data for rare diseases prevents any approach based on clinical data alone. We show, both theoretically and empirically, that our estimator always outperforms the better of the two models (expert or data) by one constant.

15:00
-
16:00
Finale Doshi-Velez
Part 1. Interpretability in Healthcare: Toward Human-based Validation and Agency

Finale Doshi-Velez
Part 1. Interpretability in Healthcare: Toward Human-based Validation and Agency

11/01/2022 - 15:00 - 16:00

There has been a lot of excitement lately about interpretable machine learning systems and explainable AI. What does it mean in the context of healthcare? In this session, we'll discuss two major ways in which interpretability can be valuable.

Human-based validation: Especially when people's health is on the line, it's crucial that we do everything we can to minimize harms and take informed risks. However, validation of AI systems for health is challenging: the measurements we collect about patients provide very partial views of that patient's health and environment, making it easy to overfit to non-causal correlations in the data. Interpretability enables human experts to inspect a system for (spheres of) correctness. I'll talk about different kinds of interpretability, the kind of questions each approach can answer, and the kinds it cannot.

Maintaining agency: In many cases, we may not want to blindly follow the output of a recommendation system. Not only may there be questions of correctness, but there may also be questions of patient preference: perhaps there are multiple reasonable options and the best choice depends on the patient's preferences; perhaps what appears to be the "best" option is really not the best given the patient's environment. I'll talk about how interpretability can help in such circumstances, as well as open questions.

16:00
-
16:30
Break

Break

11/01/2022 - 16:00 - 16:30
16:30
-
18:00
Finale Doshi-Velez
Part 2. Interpretability in Healthcare: Toward Human-based Validation and Agency

Finale Doshi-Velez
Part 2. Interpretability in Healthcare: Toward Human-based Validation and Agency

11/01/2022 - 16:30 - 18:00

There has been a lot of excitement lately about interpretable machine learning systems and explainable AI. What does it mean in the context of healthcare? In this session, we'll discuss two major ways in which interpretability can be valuable.

Human-based validation: Especially when people's health is on the line, it's crucial that we do everything we can to minimize harms and take informed risks. However, validation of AI systems for health is challenging: the measurements we collect about patients provide very partial views of that patient's health and environment, making it easy to overfit to non-causal correlations in the data. Interpretability enables human experts to inspect a system for (spheres of) correctness. I'll talk about different kinds of interpretability, the kind of questions each approach can answer, and the kinds it cannot.

Maintaining agency: In many cases, we may not want to blindly follow the output of a recommendation system. Not only may there be questions of correctness, but there may also be questions of patient preference: perhaps there are multiple reasonable options and the best choice depends on the patient's preferences; perhaps what appears to be the "best" option is really not the best given the patient's environment. I'll talk about how interpretability can help in such circumstances, as well as open questions.

Wednesday
12/01/2022
9:00
-
10:00
Ninon Burgos
Introduction to deep learning with medical imaging: from convolution to self-supervised learning (part 1)

Ninon Burgos
Introduction to deep learning with medical imaging: from convolution to self-supervised learning (part 1)

12/01/2022 - 9:00 - 10:00

In an era with increasing data and computational power, deep learning has revolutionised several computer science fields, including medicine. In particular, convolutional neural networks have been successfully applied to imaging data to perform various tasks, such as computer-aided diagnosis, image segmentation and cross-modality image synthesis.

The first part of this introduction lecture on deep learning for medical imaging will start with a short reminder on neural networks. We will then see how to exploit images as input of neural networks, detailing the convolution and the main building blocks of convolutional neural networks used for medical image classification. Finally, we will go through network structures that can be used other tasks, such as image synthesis with generative adversarial networks.

10:00
-
11:00
Julia Schnabel
Introduction to deep learning with medical imaging: from convolution to self-supervised learning (part 2)

Julia Schnabel
Introduction to deep learning with medical imaging: from convolution to self-supervised learning (part 2)

12/01/2022 - 10:00 - 11:00

The second part of the introduction plenary on machine learning in medical imaging will focus on the problem of the (lack of) quantity and quality of medical imaging data available to train machine learning models, and deep learning models in particular. After motivating this problem, we will introduce a number of concepts aimed at alleviating some of these issues. We will present for this purpose self-supervised learning using pre-text tasks, aimed at training models for seemingly irrelevant task using data without annotations. This will be followed by the problem of imbalanced classes and possible solutions. Finally, we will go through fundamental and more advanced methods for data augmentation to artificially increase the amount of training data, and will finish off with transfer learning, allowing to pretrain on one domain and retraining on a new domain with fewer data. We will use exemplars from the medical imaging community to illustrate these concepts.

11:00
-
11:30
Break

Break

12/01/2022 - 11:00 - 11:30
11:30
-
12:30
Julia Schnabel
AI-enabled medical imaging

Julia Schnabel
AI-enabled medical imaging

12/01/2022 - 11:30 - 12:30

Artificial intelligence, in particular from the class of machine / deep learning, has shown great promise for application in medical imaging. However, the success of AI-based techniques is often limited by the availability and quality of the training data. A common approach is to train methods on well annotated and curated databases of high-quality image acquisitions, which then may fail on real patient cases in a hospital setting. Another problematic is the lack of sufficient numbers of clinical label annotations in the training data, or example for early markers of disease. In this part of the plenary I will present some of our recent approaches that aim to address some of these challenges, by using AI as an enabling technique for improved image reconstruction, realistic data augmentation and further downstream tasks, using cardiac magnetic resonance imaging as an exemplar application.

12:30
-
14:00
Lunch Break

Lunch Break

12/01/2022 - 12:30 - 14:00
14:00
-
15:00
Use case - Bjoern Eskoffier
Machine Learning and Large-Scale Sensor-based Analysis for Movement Disorders

Use case - Bjoern Eskoffier
Machine Learning and Large-Scale Sensor-based Analysis for Movement Disorders

12/01/2022 - 14:00 - 15:00

Machine Learning has become a widely employed method for the analysis of large data quantities in Medicine. The number of annual Pubmed-listed publications containing the keywords “Machine Learning” is shown (as well as publications containing relevant keywords for gait analysis studies). Since the 1990s, interest in these topics is rising, making it necessary to de-mystify aspects of associated technologies, and to explain the existing opportunities and challenges to a wider audience. This was also demanded in a recent publication in Movement Disorders , where the translation of expertise from machine learning into the movement disorders community was demanded. This talk consequently aims at explaining machine learning approaches for large sensor-base data analyses in medicine, with a focus on the analysis of movement disorders.

15:00
-
16:00
Use case - Olivier Humbert & Melek Onen
Federated PET: a Federated Learning initiative in a “real world” healthcare scenario

Use case - Olivier Humbert & Melek Onen
Federated PET: a Federated Learning initiative in a “real world” healthcare scenario

12/01/2022 - 15:00 - 16:00

Among the main issues for development and clinical implementation of Deep Learning algorithms is the availability of large, curated, and representative training medical data. To this end, large scale multicentric datasets are essential to ensure generalizability of the learned model. However, the private and sensitive nature of healthcare information hampers the centralization of medical data in a unique repository. To tackle this problem, we are working on a French Federated Learning (FL) initiative for the collaborative Deep Learning analysis of 3D PET/CT images and clinical data. This FL use-case relies on Fed-BioMed, an open-source software framework, and its deployment in a real-word healthcare scenario, including a network of 10 hospitals. The interdisciplinary nature of the project requires an efficient collaboration between different actors: researchers in applied mathematics and privacy-enhancing technologies, engineers, medical doctors, informaticians of the IT System Department of hospital partners. In the presentation, we will address some of the issues of this FL project such as inter-disciplinarity communication, medical data-base collection and quality control, imaging data harmonization, development of privacy-preserving mechanisms based on cryptographic tools adapted to the federated Learning approach.

 

16:00
-
17:00
Poster session

Poster session

12/01/2022 - 16:00 - 17:00
17:00
-
17:15
Break

Break

12/01/2022 - 17:00 - 17:15
17:15
-
18:15
Roundtable - FREE ACCESS

Roundtable - FREE ACCESS

12/01/2022 - 17:15 - 18:15

The Roundtable, with personalities from academia, industry and public institutions, will address the theme “Data-driven approaches in healthcare and precision medicine: Challenges & opportunities” and include:

- Raphaël Porcher (animator), Professor of Biostatistics at Université de Paris, Co-director of the Centre Virchow-Villermé Paris Berlin,Member of the METHODS team of CRESS-UMR1153 and Chair at the PR[AI]RIE Artificial Intelligence Institute

- James J. Collins, Termeer Professor of Engineering and Medical Sciences, Broad Institute of MIT and Harvard and The Wyss Institute,

- Kit Roes, Professor of Biostatistics at Radboud University Medical Center Nijmegen and Visiting Professor of Clinical Trial Methodology at University Medical Center Utrecht)

- Christian Ohmann, Head of the Clinical Trial Coordination Center (KKS) at the Medical Faculty of the Heinrich Heine University in Düsseldorf, Germany,

- Billy Amzal, CEO Quinten Health,

- Mohamed-Ramzi Temanni, Scientific Director, Head of France AI/Genomics, Computational Sciences, Janssen R&D - Global Development.

18:15
-
18:30
Conclusion & perspectives

Conclusion & perspectives

12/01/2022 - 18:15 - 18:30
Thursday
13/01/2022
9:00
-
10:30
Practical sessions

Practical sessions

13/01/2022 - 9:00 - 10:30
10:30
-
11:00
Break

Break

13/01/2022 - 10:30 - 11:00
11:00
-
12:30
Practical sessions

Practical sessions

13/01/2022 - 11:00 - 12:30
12:30
-
14:00
Lunch break

Lunch break

13/01/2022 - 12:30 - 14:00
14:00
-
15:30
Practical sessions

Practical sessions

13/01/2022 - 14:00 - 15:30
15:30
-
16:00
Break

Break

13/01/2022 - 15:30 - 16:00
16:00
-
17:30
Practical sessions

Practical sessions

13/01/2022 - 16:00 - 17:30
Friday
14/01/2022
9:00
-
10:30
Practical sessions

Practical sessions

14/01/2022 - 9:00 - 10:30
10:30
-
11:00
Break

Break

14/01/2022 - 10:30 - 11:00
11:00
-
12:30
Practical sessions

Practical sessions

14/01/2022 - 11:00 - 12:30
12:30
-
14:00
Lunch break

Lunch break

14/01/2022 - 12:30 - 14:00
14:00
-
15:30
Practical sessions

Practical sessions

14/01/2022 - 14:00 - 15:30
15:30
-
16:00
Break

Break

14/01/2022 - 15:30 - 16:00
16:00
-
17:30
Practical sessions

Practical sessions

14/01/2022 - 16:00 - 17:30