Evaluating Algorithms as Health Policies: Microsimulation to Assess Health Effects Before Deployment

1 credits · Graduate

Predictive medical AI models are often imperfectly accurate and work better for some subpopulations than others. A range of methods exist for understanding how they will affect existing health inequities, from algorithm-centric performance evaluations, or audits on one end, and setting-centric implementation evaluations on the other. In this course, students will be introduced to a third path, drawn from the field of decision science. We will study microsimulation modeling, in which the effect of using an algorithm - given different scenarios and assumptions about potential harms and benefits - can be simulated computationally. Students will be introduced to Markov and microsimulation models, explore the process of generating and formalizing assumptions about model effectiveness, and reason about the appropriateness of these evaluations in different settings. The course is designed for researchers building and evaluating medical algorithms, as well as clinicians considering their own use of predictive tools in clinical care. The course will include an optional python-based coding project component.