Temporal Inverse Probability Weighting for Causal Discovery in Controlled Before–After Studies: Discovering ADEs in Generics
Adverse drug events (ADEs) cost society lives and an estimated $30 billion per year in the USA alone. Their prevalence has led to the public losing trust in the safety of drugs, especially generics (e.g., Eban, 2019). These concerns have motivated the wide study of methods for general ADE discovery, but discovering ADEs in generic drugs challenges causal discovery methods with a scenario of multiple treatments over time, a scenario which presents new problems and opportunities for machine learning. In response, this research develops methods for causal discovery based on analyzing controlled before–after studies with differential prediction and temporal inverse probability weighting. These methods are easy to realize by employing off-the-shelf machine learning classifiers. Experiments on both synthetic and real electronic health records demonstrate the ability of the methods to control for confounding, discover generic-specific ADEs in synthetic data, and hypothesize brand–generic differences in real-world data that agree with known ones. These are the abilities that causal discovery methods need for helping establish the facts of generic drug safety.
