Abstract

The volumes of medical data being recorded are now far beyond what human experts can analyze, especially at the bedside. At the same time, a lot of essential information goes unrecorded; any purely machine system will quickly hit fundamental statistical limitations. Thus, to make a positive impact in healthcare, we need robust tools that can communicate their limitations and assumptions to the ultimate human decision-maker.

Whether it is building models, optimizing treatment policies, or validating outputs, my lab addresses this challenge by keeping human-machine validation and integration in mind from the start. In this talk, I will discuss how our pursuit of small, inspectable models lead us to find (and resolve) a decade-old failing of supervised generative models; how those models (seemingly) resulted in better strategies for managing hypotension in the ICU ; how we realized that some of their strategies were (perhaps) bogus and how we regained (some of) our trust via novel ways of incorporating human input into statistical off-policy evaluation; and how our recommendations for interpretability and careful validation are being heard throughout the world. All through the way, I will emphasize the many interesting technical questions whose answers have potential for real impacts in health.