Interpretable (not just posthoc-explainable) medical claims modeling for discharge placement to reduce preventable all-cause readmissions or death

Chang TL, Xia H, Mahajan S, Mahajan R, Maisog J, et al. (2024). PLOS ONE 19(5): e0302871. https://doi.org/10.1371/journal.pone.0302871

Overview

Hospital readmissions remain a major cost and quality concern, yet most predictive models are black boxes that clinicians cannot meaningfully interrogate. This paper presents an inherently interpretable multilevel Bayesian framework that mimics the piecewise linearity of ReLU-activated deep neural networks while remaining fully transparent. The model uses medical claims data to predict hospital readmission and death, with a focus on discharge placement decisions and causal estimation of local average treatment effects adjusted for confounding.

Trained on a 5% sample of Medicare beneficiaries (2008--2011 inpatient episodes, tested on 2012), the model achieves an AUROC of approximately 0.76 on predicting all-cause readmissions or death within 30 days of discharge — competitive with XGBoost and Bayesian deep neural networks. Unlike those approaches, this model provides exact global interpretations: identifying relative risk factors and quantifying the effect of discharge placement. The paper also demonstrates that the posthoc explainer SHAP fails to provide accurate explanations for this class of models, underscoring the value of built-in interpretability over post-hoc approximation.