Abstract: For complex diseases like depression, choosing a successful treatment from several possible drugs remains a trial-and-error process in current clinical practice. By applying statistical machine learning to the electronic health records of thousands of patients, can we discover subtypes of disease which both improve population-wide understanding and improve patient-specific drug recommendations? One popular approach is to represent noisy, high-dimensional health records as mixtures of low-dimensional subtypes via a probabilistic topic model. I will introduce this common dimensionality reduction method and explain how off-the-shelf topic models are misspecified for downstream prediction tasks across many domains from text analysis to healthcare. To overcome these poor predictions, I will introduce a new framework -- prediction-constrained training -- which learns interpretable topic models that offer competitive drug recommendations. I will also discuss open challenges in using machine learning to improve clinical decision-making. 

Michael Hughes is an Assistant Professor of Computer Science at Tufts University.