Whose Readmission Gets Missed? A Fairness and Calibration Audit of Hospital Readmission Prediction
Readmission prediction models are usually reported with a single number — one AUC, computed over everyone. That number can be accurate and still hide the only question a clinician or a patient actually cares about: for whom does the model fail? We audit, rather than merely train, a 30-day readmission predictor on the UCI Diabetes 130-US Hospitals dataset. After first-encounter-per-patient filtering and removal of patients discharged dead or to hospice, 69,973 patients remain, of whom 9.0% were readmitted within 30 days. A gradient-boosting model reaches a test AUC of 0.636 and a logistic regression 0.624 — unremarkable, and consistent with the published ceiling for administrative readmission models. The audit is where the information is. At the software-default threshold of 0.5 the gradient-boosting model identifies zero of the 943 readmitted test patients. At a shared, validation-calibrated operating threshold, recall varies systematically: 0.41–0.82 across age bands, 0.48 for AfricanAmerican patients versus 0.65 for Caucasian patients (bootstrap confidence intervals do not overlap), and 0.58 versus 0.64 by gender — while per-group calibration error stays small everywhere (ECE ≤ 0.041). The failure is therefore not mainly miscalibration; it is the distribution of thresholded errors. We then test the standard fix, per-group thresholds fitted on validation data. It works where groups are large — the age recall gap falls from 0.404 to 0.150 and the gender gap from 0.065 to 0.016, at a cost of about one point of precision — but for race it does not close the headline gap, and in its ungated form it harms the smallest groups: the Hispanic group's recall falls from 0.556 to 0.333 because its threshold was estimated from a handful of validation positives. A support-gated variant narrows the AfricanAmerican–Caucasian gap from 0.166 to 0.099 at negligible overall cost. Our conclusion is procedural: a fairness audit that reports gaps without confidence intervals, and a mitigation applied without a minimum-support rule, can both mislead — the groups most at risk of being missed are precisely the groups too small for naive post-processing to reach.
Authors
- Saagar Sengupta
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-13
- DOI
- https://doi.org/10.5281/zenodo.24208995
- Primary Topic
- Machine Learning in Healthcare
- Type
- preprint