Discovering complex counterfactual phenotypes in the presence of inconclusive outcomes

Randomized trials are frequently inconclusive in determining the most appropriate interventions for a given cohort of patients. In this project, we will try to discover if there are subcohorts of individuals that benefit from interventions in terms of all potentially relevant outcomes. We will try to train models to determine if patients benefit across multiple outcomes and, if yes, what are the covariates that correspond to patients benefiting from them. We will then experiment with latent variable methods that can simultaneously recover actionable phenotypes/subgroups from a clinical-making standpoint. We will work with mid-sized NIH tabular datasets (ALLHAT/ACCORD/SPRINT/BARI2D) involving measurements of patient characteristics, medications, and vitals collected at multiple time points. The data also includes time to event outcomes like time to stroke, heart failure, mortality, etc.