A Two-Stage Prescriptive Equity Audit for Clinical AI Recourse: Feasibility, Burden Decomposition, and Temporal Validation

This repository contains the official replication package, source code, and data specifications for the study: **"A Two-Stage Prescriptive Equity Audit for Clinical AI Recourse: Feasibility, Burden Decomposition, and Temporal Validation"** ### Abstract**Objective:** To evaluate whether a clinically constrained algorithmic recourse audit produces reproducible and temporally stable prescriptive equity signals.**Materials and Methods:** We trained an Explainable Boosting Machine (EBM) predicting severe tooth loss using 2022 Behavioral Risk Factor Surveillance System (BRFSS) data (N=433,772). High-risk adults entered a constrained recourse search enforcing immutability locks, behavioral directionality, physiological safety floors, and sparsity (k≤3). We calculated Social Indicators of Disparity Index (SIDI)-neutral model-movement costs—using SIDI strictly for stratification, not cost inflation—and applied Oaxaca-Blinder (OB) decomposition among reachable adults. Reproducibility and temporal transport were validated using a 2022 holdout and frozen 2024 BRFSS (N=448,213) without retraining.**Results:** The locked 2022 model transferred to 2024 with preserved discrimination (weighted AUC: 0.855→0.858; Brier: 0.088→0.086; calibration slope: 1.014). Constrained reachability remained stable but low (2022 full: 10.6%; holdout: 10.5%; 2024 transfer: 9.8%). Primary SIDI-neutral unexplained OB residuals were near-null and tightly bounded (0.013 to 0.005; all 95% CIs crossing zero). Conversely, unexplained income-axis residuals persisted (0.108 to 0.045) alongside a four-fold reachability gap. ### Repository ContentsThe uploaded archive contains the components necessary to reproduce the empirical results, figures, and tables presented in the manuscript:* `src/`: Core Python/R scripts for training the Explainable Boosting Machine (EBM), generating clinically constrained algorithmic recourse, and executing the two-stage Oaxaca-Blinder equity decomposition.* `data/`: Derived and preprocessed cohorts extracted from the 2022 and 2024 Behavioral Risk Factor Surveillance System (BRFSS) databases.* `models/`: Frozen and locked model specifications used during the temporal validation phase.* `README.md`: Detailed documentation including software dependencies, environment setup, and instructions for execution. ### Usage and CitationPlease cite the corresponding peer-reviewed journal publication when utilizing any code or assets from this repository.