Leveraging Operations Research for Responsible AI in Medicine

An estimated 133 million - or nearly half - of all Americans suffer from one or more chronic diseases. Chronic diseases consistently account for 5 of the top 10 leading causes of death. Moreover, 90% of annual healthcare expenditures are attributed to chronic diseases. Atherosclerotic Cardiovascular Disease (ASCVD) and Type 2 Diabetes (T2D) specifically comprise a significant portion of these chronic diseases in the United States, and advances in disease management for ASCVD and T2D can potentially reduce healthcare expenditures and adverse outcomes associated with these diseases. To this end, many algorithmic approaches to disease management have been proposed in the name of personalized medicine. Yet, few have accounted for two key principles in the emerging area of Responsible AI - namely interpretability and equity. In this talk, I will discuss my recent work on designing models to facilitate disease management for ASCVD and T2D with a focus on interpretability and equity. Chiefly, this work includes the formulation and analysis of interpretable treatment planning for Markov Decision Processes (MDPs). We specifically analyze the problems of optimizing monotone policies, which increase treatment intensity with worsening patient health, and optimizing class-ordered monotone policies (CMPs), which generalize monotone policies by imposing monotonicity over classes (i.e., groups) of states and actions. We establish key analytical properties of both problems and propose exact formulations for optimizing interpretable policies in general. Next, we define and analyze the price of interpretability (PI), proving that the CMP’s PI does not exceed the PI of the monotone policy. We then design and evaluate MDPs for hypertension treatment planning using a nationally representative dataset of the United States’ population. At the patient level, we find that optimal MDP-based policies may be unintuitive, recommending more aggressive treatment for healthier patients than sicker patients. Conversely, monotone policies and CMPs never de-escalate treatment, reflecting clinical intuition. Across 66.5 million patients, optimized monotone policies and CMPs save over 3,246 quality-adjusted life years per 100,000 patients over current clinical guidelines, while paying low PIs. We conclude that interpretable policies can be tractably optimized, drastically outperform existing guidelines, and pay low PIs - potentially increasing the acceptability of decision-analytic approaches in practice. To conclude the talk, I will briefly discuss ongoing works that touch on interpretable personalized treatment via counterfactual optimization and a Responsible AI-based framework to design and evaluate risk estimation models for ASCVD and T2D.

Speaker: Gian-Gabriel Garcia, Georgia Institute of Technology

Room 3108