WP4 Proof-of-concept projects

To improve characterization of health trajectories with patients’ risk stratifications and prediction of responses to interventions, we will include the scientific knowledge gained from WP2 and used the secured data platform established in WP3 to: 1) identify specific patient profiles or typical longitudinal patterns of evolution over time (trajectories), and 2) build novel predictive models and tools useful for clinical management that take into account the heterogeneity of patients’ trajectories.

The construction of patient trajectories from a large set of multidimensional covariates from different sources and repeatedly measured over time raises methodological challenges that will be addressed by a dedicated biostatistical team, combining the strengths of several biostatisticians (HP2, LIG, UGA data institute) in partnership with national and international experts.

We will develop strategies based on Exploratory Data Analysis (EDA) that empowers medical professionals with the ability to test multiple hypotheses on patient cohorts in an iterative, exploratory manner. We will provide and evaluate three EDA approaches: manual where the medical experts choose the next exploration step; partially-guided where the system suggests the next hypothesis to be tested but the medical expert is allowed to override the suggestion; and fully-guided where the medical expert acts as an observer and the system runs a series of hypotheses on an input cohort. These exploration modes will be made possible via learning exploration policies using Deep Reinforcement Learning, a Machine Learning method that simulates an agent to recommend the best sequence of hypotheses to be tested on a cohort. This will require the design of reward mechanisms in concertation between data scientists and medical experts.