My research is multidisciplinary, with a strong focus on biomedical applications. I tackled challenges in different modalities such as computer vision, tabular data, and token sequences of clinical data, specifically when they are scarce, high-dimensional or noisy. My main tools are deep and statistical learning.

Characterization of clinical trajectories 2025 – now

At CBIO (Mines Paris – PSL & Institut Curie), I work with Eloïse Berson on longitudinal electronic health records (EHR). I build deep sequential models (GRU, transformer, neural ODEs) that predict patient survival from sparse, irregular follow-up data, and study how foundation models can help design better patient cohorts for biomedical discovery. The main focus of the project is to then be able to characterize patients trajectories for long-term illnesses such as cognitive decline (e.g., Alzheimer’s disease), with also a parallel interest in mid-term cancer such as breast, lung, pancreatic and brain cancers. Since 2026 this work is funded by a PR[AI]RIE AI Fellowship.

Accelerating numerical simulations with AI 2024

While the start-up NPCo (ex-Augur) was being developed at Inria Start-Up Studio incubator, with Emmanuel Menier and Matthieu Nastorg, I benchmarked deep learning surrogates for numerical simulations on 2D and 3D meshes. These models predict physical fields (e.g., temperature, pressure) and physical scalar values (e.g., drag coefficients) much faster than classical solvers, to speed up the design of wind turbines, rotors or medical devices for example.

Generative models for transcriptomics PhD, 2021 – 2024

Deep learning requires many training samples, yet gene expression cohorts are small (a few hundred patients for ~20,000 genes in bulk-RNA data), leading to a small n, large p problem. During my PhD with Blaise Hanczar (IBISC) and Michèle Sebag (Inria TAU), I used deep generative models for data augmentation to improve cancer phenotype prediction from RNA-seq and microarray data. I benchmarked VAEs, GANs and diffusion models for transcriptomics, with adapted quality metrics (Bioinformatics 2023, BMC Bioinformatics 2026). I also proposed GMDA, a frugal, non-adversarial generative model for tabular data (ECML-PKDD 2024).

Covid-19 forecasting 2020 – 2021

As a finalist of the XPRIZE Pandemic Response Challenge, I worked on LSTM models to forecast infection rates across +40 countries. I then worked as an intern with Isabelle Guyon at Inria to combine epidemiological compartmental models (SIR) with machine-learning confidence intervals (ML-CI) built from random forests.

Deep learning for astrophysics 2019

My first research internship experience, with Samuel Farrens at Cosmostat (CEA), was about using transfer learning (a fine-tuned VGG-16) to detect blended galaxies in simulated survey images (A&A 2022).