Description du sujet de thèse

The objective of this thesis is to develop dynamic AI models predicting cardiac arrest (CA) risk by exploiting individual medical trajectories from structured temporal data (SNDS and CEMS). The algorithms developed will integrate the longitudinal dimension of care pathways.
Using advanced machine learning methods adapted to longitudinal data (EHR), the project will analyze sequences of diagnoses, medical procedures, and prescriptions to accurately model the evolution of risk over time. Vector representations (embeddings) will be explored to capture the complex interactions between clinical events, particularly using techniques such as Time2Vec (to encode continuous temporal dimensions) or FAN (Fourier Analysis Networks) to capture frequency patterns.
To cope with the large volume of data (long sequences, numerous patients), Transformer architectures will be explored. In particular, the Performer (FAVOR+) model, which is based on a kernelized attention approximation, will efficiently process long clinical sequences thanks to linear complexity, while retaining the ability to model distant dependency relationships.

In addition to Transformer-type or more traditional boosting approaches, several complementary avenues could be explored to enrich risk modeling: Hierarchical Bayesian methods, Temporal stochastic process models, Hawkes processes.

Contexte de travail

This thesis will be supervised by E. Bacry (CEREMADE, Université de Paris-Dauphine, Institut PR[AI]RIE) and co-supervised by Prof. Xavier Jouven (APHP, INSERM, Institut PR[AI]RIE). Most of the work will be done within the INSERM team (unit U970) 56 rue Leblanc 75005 Paris.