A HYBRID BIAS-MITIGATION FRAMEWORK FOR FAIR CLINICAL READMISSION PREDICTION
DOI:
https://doi.org/10.26577/bcdn8166Keywords:
algorithmic fairness, responsible artificial intelligence, clinical prediction, bias mitigation, equalized odds.Abstract
Machine-learning models increasingly guide the allocation of clinical care, yet models optimised for aggregate accuracy can encode substantial disparities in error rates across demographic groups. This study addresses fair prediction of 30-day hospital readmission for diabetic inpatients under two constraints that are common in practice but rarely modelled together: the deployed model must not use a protected attribute at decision time, and the positive class is rare. We propose FAIR-Health, a hybrid bias-mitigation framework that couples Kamiran Calders instance reweighing (pre-processing) with a differentiable equalized-odds surrogate penalty embedded in the training objective (in-processing). The result is a single attribute-blind model with one tunable control over the fairness/accuracy trade-off. Using the UCI Diabetes 130-US Hospitals cohort (71,518 patients after leakage-free, patient-level deduplication), we benchmark FAIR-Health against pre-, in-, and post-processing baselines across ten random data splits (seeds) with paired significance testing. Compared to the unmitigated gradient-boosted model, FAIR-Health reduced the equalized-odds difference by 44.7% for gender and 65.6% for race, at a within-model-family AUROC cost of at most 0.004 and with no loss of calibration. Measured against the unmitigated logistic model, the true-positive-rate gap between racial groups fell from 0.111 to 0.019. We further show that hard-constraint methods attain near-zero unfairness only by collapsing to a near-constant classifier, which removes their main practical value. The framework and benchmark give practical guidance for deploying equitable clinical prediction models.











