PhD supervisor : Chantal Soulé-Dupuy – IRIT, Toulouse
PhD co-supervisor : Paul Monsarrat – Restore, Toulouse
Aging is an overly complex biological process, involving multiple mechanisms at different levels, from molecular to tissue scale. With the increase of life expectancy, the challenge is to predict the physiological age to reach healthy aging, namely the one’s ability to adapt and efﬁciently respond to stressors. The recent use of machine learning (ML) strategies may accurately model the physiological age to prevent age-related disruption. However, black boxes properties of most ML algorithms do not support the understanding of their internal decision-making mechanisms (i.e., the explainability), crucial to highlight the critical variables on the prediction.
Using a strategy based on the coalition of attributes we recently develop algorithmic solution, whose performances and prediction qualities are superior to the gold standard SHAP (Shapley Additive exPlanation) approach. We propose to consider the prediction explanations, not as a final objective, but as a new data space allowing to impulse a better interaction between the ML model and the end user. This thesis project will focus on the development of an original framework for the biomedical community to explore in depth predictions of health evolution with age using an ML model. It will allow the detection of subpopulations with specific predictive explanations, organizing them hierarchically to ensure easy exploration. As a proof of concept, we will apply this framework to physiological age (in-house database of >40,000 subjects, 200 variables), to better understand biological determinants of ageing, their interactions, putative causal chains, and underlying physio-pathological mechanisms.
Key words: aging, machine learning