KAISEN: Reproducible Subgroup Fairness Auditing for Clinical Risk Models

Clinical risk models often maintain high aggregate performance while masking significant error rate disparities across patient subgroups. To address this, the authors developed KAISEN, a five-phase audit pipeline designed to stress-test fairness components including stratification, disparity measurement, mechanism diagnostics, post-hoc mitigation, and drift monitoring. By using a synthetic benchmark of 16 disease tasks and 15 social-determinant axes, the researchers evaluated where these audit components fail and whether those failures are detectable in the audit output.

The study highlights four critical insights regarding fairness auditing. First, statistical significance in fairness metrics is highly sensitive to the minimum detectable effect; standardizing equalized-odds differences (EOD) by this floor significantly improves the correlation between significance counts and actual disparities. Second, while per-group threshold optimization consistently reduces EOD, group-wise Platt scaling—often considered a superior calibrator—acts unpredictably, suggesting that auditors should report the variance of these interventions rather than just the mean effect.

Third, the mechanism diagnostic, intended to identify the source of disparities, fails silently when proxy columns are misspecified, providing no signal that the diagnostic itself is unreliable. Finally, longitudinal monitoring using CUSUM (Cumulative Sum) charts shows that false alarms and detection failures are driven more by cohort-specific characteristics than by the disease itself, meaning that alarm thresholds tuned on one population do not reliably transfer to others.

This work challenges the assumption that audit pipelines are neutral, objective instruments. By demonstrating that common diagnostic and mitigation tools can fail without warning, the authors emphasize that clinical AI audits must be stress-tested against known ground truths. The study provides a framework for institutions to evaluate which fairness metrics are most consequential for their specific clinical decision contexts—such as sepsis detection versus hospital readmission—ensuring that audits are not just performed, but are actually informative and actionable.