Interpretable (not just posthoc-explainable) heterogeneous survivor bias-corrected treatment effects for assignment of postdischarge interventions to prevent readmissions

Xia H, Chang JC, Nowak S, Mahajan S, Mahajan R, Chang TL, Chow CC (2023). Proceedings of the 8th Machine Learning for Healthcare Conference, PMLR 219:884-905.

Overview

Presented at MLHC 2023, this paper tackles the problem of assigning postdischarge interventions to prevent hospital readmissions. A key challenge in this setting is survivor bias: patients who are readmitted early never receive the full course of an intervention, which can distort estimates of treatment effectiveness.

The paper introduces an interpretable Bayesian approach that estimates heterogeneous treatment effects while correcting for survivor bias. Rather than relying on black-box models with posthoc explanations, the method produces inherently interpretable estimates of which patients benefit most from specific postdischarge interventions. This allows clinicians to make individualized, evidence-based decisions about intervention assignment at the point of discharge, with full transparency into the model's reasoning.