Edward De Brouwer

Research

Machine learning modeling of time-dependent patient trajectories

We are now in the era of systematic collection and massive storage of data. The biomedical sector is inline with this trend and is an important source of rich data. Specifically, patients records are now stored electronically and contain increasingly comprehensive information. This availability opens up perspectives of significant advances in the medical field. It will allow assessing patient profiles more accurately and lead to desirable outcomes: (1) personalized healthcare and enhanced diagnostics, (2) better selection of participants in clinical trials to improve the efficiency of those trials, (3) a deeper understanding of relations between pathologies, drugs and genomic information. However, historical patient records are challenging to handle, because of intrinsic properties such as time-dependence or sporadic measurements.

This research aims at developing robust machine learning models to retrieve most relevant information from this data. To overcome the hurdles previously mentioned, we propose two innovative approaches. Both rely on the modeling of hidden variables that represent an unobserved disease state. One is matrix-factorisation with dynamic latent vectors, the other is the use of recurrent neural networks (LSTMs). More than extending the current state of the art implementations, they provide a new perspective on the problem. To make the most of these, we will also provide a scheme to share data and our models across distinct entities without having to disclose any information. In this way, we address the pressing issue of data sharing and centralization.

Publications