Auditing Algorithmic Fairness in Healthcare: An Empirical Test of the AIMS Governance Framework on 99,492 Diabetes Readmissions

This repository contains the code, data outputs and figures source for the empirical evaluation of the AIMS Governance Framework (AI Management System) applied to algorithmic fairness auditing in hospital readmission prediction. The paper makes three empirical contributions: 1. ML fairness audit: A Random Forest classifier is trained on the public UCI Diabetes 130-US Hospitals dataset (N=99,492 encounters; target: 30-day readmission). The AIMS Pillar 2 algorithmic impact assessment identifies three utilization-proxy features (`number_inpatient`, `number_emergency`, `number_outpatient`) that inflate headline AUC by 0.050 units (0.659 → 0.609, bootstrap 95% CIs non-overlapping; cross-seed robust Δ = 0.050 ± 0.0005 over 5 seeds). 2. Subgroup fairness analysis: AUC is stratified by race, gender, and age subgroups, with subgroup-specific bootstrap 95% CIs. The apparent fairness improvement from proxy removal (mean gap 0.125 → 0.086) is reported as **not statistically significant** (bootstrap CI on the change includes zero, [–0.056, +0.062]); the African-American AUC drops from 0.628 to 0.562 under proxy removal, a subgroup-specific regression. 3. Agent-based IFP simulation: The Institutional Feedback Processing (IFP) stability criterion is tested computationally across 60 simulated organizations × 12 PDCA cycles × 30 seeds. IFP reduces aggregate governance signal load by 78.8% versus a no-IFP control.