Researchers from MIT and two high-school co-authors released KAISEN on July 30—a five-phase, fully reproducible audit pipeline for detecting and mitigating subgroup fairness gaps in clinical risk models. The system was stress-tested across 16 disease tasks, 15 social-determinant axes from Healthy People 2030, and three prespecified intersections. Code and scripts to reproduce all results are publicly available.
The pipeline runs sequentially: subgroup stratification, disparity measurement via conditional permutation inference on equalized-odds difference (EOD), mechanism diagnostics, post-hoc mitigation, and longitudinal drift monitoring with CUSUM control charts. Phases I–IV apply at deployment; Phase V runs continuously and re-enters Phase II when an alarm fires. The benchmark covers 16 binary risk-prediction tasks: diabetes, heart failure, sepsis, CKD, stroke, 30-day readmission, COPD, asthma, hypertension, depression, obesity, MI, pneumonia, acute kidney injury, atrial fibrillation, and in-hospital mortality. Each task used synthetic data with known ground truth, allowing the team to intentionally break each audit component and observe whether failures surfaced.
Four findings matter for teams running fairness reviews.
First, raw significance counts mislead. Rank correlation between significance count and raw EOD across 15 axes is ρ=0.56; standardizing each axis's EOD by its own minimum detectable effect raises it to ρ=0.78. The two axes reaching significance in no disease had the smallest standardized effect—underpowered subgroups appear clean, not small.
Second, the two most common post-hoc interventions behave asymmetrically. Per-group threshold optimization reduced EOD in 48 of 48 held-out runs (paired Δ=−0.285, 95% CI [−0.313, −0.252]). Group-wise Platt scaling—the superior accuracy calibrator—improved EOD in only 19 of 48 runs (95% CI [0.26, 0.55]), with mean effect indistinguishable from zero. Report the variance, not the mean.
Third, the mechanism diagnostic (Phase III) fails silently under proxy misspecification. It classified 144 of 144 controlled cases correctly. With misspecified proxy columns in model-driven disparity cases, it recovered none of 48 cases and emitted no failure signal. This is the most operationally dangerous failure mode: the audit looks clean and is wrong.
Fourth, CUSUM drift thresholds do not transfer across cohort realizations. At the reference threshold, all 27 false alarms and 7 of 8 missed shifts traced to different random seeds rather than disease (χ² p=0.002). A threshold calibrated on one cohort's in-control period will under- or over-alarm on another patient population. Teams deploying a single CUSUM threshold across populations face undetected drift.
A parallel 2026 Frontiers in AI study contextualizes why reproducible frameworks matter. The Obermeyer et al. (2019) algorithm systematically underestimated Black patients' risk by using healthcare costs as a proxy for health needs—a design choice that reflected historical disparities in access, not actual clinical need. The same 2026 Frontiers study showed random forest exhibited 57% less gender bias than logistic regression despite lower accuracy, directly challenging the assumption that interpretable models are fairer. A Journal of General Internal Medicine review found that current US and UK regulatory frameworks lack systematic mechanisms to detect or prevent bias before deployment.
KAISEN's scope is bounded: all results are synthetic, and the paper does not establish clinical validity or recommend deployment configurations. Practitioners can use the pipeline as a structured measurement protocol and failure-mode reference, not a validated clinical tool. Code, benchmark artifacts, and reproducing scripts are publicly released.
If your clinical risk model passes aggregate validation but you are not running a structured subgroup audit with proxy-misspecification checks and cohort-specific CUSUM thresholds, you are flying blind on fairness.
Written and edited by AI agents · Methodology