Health and social sciences

Electronic health records

Electronic health records are typically data collected by the care system. As a result, they cover a large number of individuals over a long period of time. However, they are observational data: interventions, such as drug prescription, have not been done according to a systematic experimental plan. As such, they raise many causal inference challenges (which we study in link with machine learning). In addition, they are complex relational data, structured for clinical practices or accounting, and not research. Analyzing them faces challenges of missing values, as well as statistical modeling across multiple sources of relational data with varying representations. The databases are full of non normalized snippets of text, for which we develop dedicated NLP approaches.

We apply our statistical learning research on large electronic health records (in partnership with APHP, HAS, and on the French healthcare system: SNDS). The related questions are that of personalized medicine (individualized prediction models) as well as treatment efficacy, for instance to answer public-health questions.

Epidemiological reasoning

From a health-research perspective, the questions are typical of epidemiology and draw from the corresponding line of thoughts: all kind of biases must be accounted for, such as a sampling bias between the study population and the target. Where we innovate is that we strive to replace the traditional parametric models with non-parametric estimates built from machine-learning models.

AP-HP’s “entrepot de données de santé”

We have a tight collaboration with AP-HP, the Paris hospitals, that unite 39 different hospital in a consistent database covering 10 million patients a year. This collaboration was born at the beginning of the COVID crisis, where we used the corresponding database (known as the EDS, “entrepôt de données de santé”) to follow the unfolding of a new disease, building on the fly prognosis models and evaluating treatment efficiency.

Educational data mining, learning analytics

Learning platforms collect a large amount of data that can be used to personalize learning. Educational data mining is about developing mathematical models of student learning (knowledge tracing, spaced repetition systems), for example students attempting programming exercises. Learning analytics is focused on representing information for teacher decision-making (e.g. dashboards in MOOCs).

Related publications

Medical applications

Education-related

COVID-related