Abstract

This presentation will showcase a clinical application of state-of-the-art models for pulmonary embolism (PE) risk stratification. We studied whether models can accurately classify risk from only medical records and biomarkers extracted from CTPA images. We benchmarked different approaches to combine global features with rich pulmonary vascular information; we added vascular biomarkers to tabular models and applied graph neural networks (GNNs) on the vascular tree's intrinsic graph representation. On a private dataset (n=353) with uniquely complete data for PE risk stratification, we showed counter-intuitive results that neither vascular biomarkers nor full vascular graphs improve performance over simpler clinical biomarkers. Finally, we investigated hypotheses, on both models and data, that could explain this suboptimal performance.

Bio

Nathan Painchaud obtained a joint PhD in computer science and signal and image processing from the Université de Sherbrooke, Canada, and INSA Lyon, France, in 2024. He is now finishing a 2-year postdoc at the CREATIS laboratory (INSA Lyon). On the technical side, his research interests are in representation learning and combining multi-modal data. On the applicative side, he uses these methods on healthcare data to solve clinical problems like population analysis and risk stratification.

Additional Information

The seminar will be given in 🇫🇷 French 🇫🇷, and is open to everyone, either in person in room 333 (Lyon campus) or online via Teams.