David Sontag (MIT) - Learning Deep Markov Models for Precision Medicine

Colloquium online lecture by David Sontag (MIT), Jun 7, 2021, at The School of Computer Science and Engineering, The Hebrew University.


I present a new approach to learning from temporal data, coupling deep learning with probabilistic inference. Applied to learning disease progression models from clinical data, our algorithms learn rich representations that are capable of answering counterfactual questions such as which treatment is most appropriate to which patient, providing a new theoretical framework for precision medicine. Making valid causal inferences from observational data requires a number of assumptions to be satisfied. I show how machine learning can be used to test and explain one of these (overlap) and how machine learning can help circumvent another (hidden confounding). Along the way, I'll make connections to recent work on domain adaptation and dataset shift. Finally, I discuss my vision for the future, where these methods are scalably used to guide millions of patients' health care. Doing so will require policy and legislative changes to improve health data collection and curation, new algorithms for extracting treatment and outcomes from clinical text, and advances in human-computer interaction to safely and effectively explain algorithm predictions to patients and providers.