KAISEN: Reproducible Subgroup Fairness Auditing for Clinical Risk Models

Overview§

KAISEN is a reproducible five-phase audit pipeline for subgroup fairness in clinical risk models: (1) subgroup stratification, (2) disparity measurement, (3) mechanism diagnostics, (4) post-hoc mitigation, and (5) drift monitoring. The authors evaluate each component to the point of failure using a synthetic benchmark with known ground truth, covering 16 disease tasks, 15 social-determinant axes (from Healthy People 2030), and 3 prespecified intersections.

Core Methodology§

The pipeline is modular and stress-tested independently. Two key components are analyzed in depth:

• Disparity measurement: Uses the equalized-odds difference (EOD), defined as the maximum absolute difference in false-positive rate (FPR) and false-negative rate (FNR) across groups:

$$\text{EOD} = \max_{g \in \mathcal{G}} \left( |\text{FPR}_g - \text{FPR}_{\text{ref}}|,\; |\text{FNR}_g - \text{FNR}_{\text{ref}}| \right)$$ The authors show that statistical significance counts correlate only weakly with raw EOD ($\rho = 0.56$), but strongly once EOD is standardized by the axis-specific minimum detectable effect (MDE): $\rho = 0.78$.

• Post-hoc mitigation: Compares per-group threshold optimization against group-wise Platt scaling. Threshold optimization finds a decision threshold $t_g$ per group minimizing EOD, while Platt scaling fits a logistic calibration map $p_g = \sigma(\alpha_g s + \beta_g)$. Results: threshold optimization reduces EOD in 48/48 held-out runs (paired delta = -0.285, 95% CI [-0.313, -0.252]), while Platt scaling improves EOD in only 19/48 runs with mean effect near zero. The paper argues the correct metric is variance, not the average.

• Mechanism diagnostics: A classifier designed to identify whether disparities stem from proxy variables or genuine signal. It achieves 144/144 accuracy on controlled cases but 0/48 on model-driven cases under proxy misspecification, with no internal signal indicating failure.

• Drift monitoring: Uses CUSUM control charts for temporal drift. False alarms and missed shifts are driven more by cohort realization (seed variation) than by disease type: at the reference threshold, all 27 false alarms and 7/8 missed shifts come from different seeds (chi-squared $p = 0.002$), implying thresholds do not transfer across cohorts.

Implementation Sketch§

The following simplified Python code illustrates the core mitigation evaluation loop (threshold optimization vs. Platt scaling):

import numpy as np
from sklearn.calibration import CalibratedClassifierCV
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import roc_curve

def threshold_optimize(probs_all, y_all, groups, ref_group=0):
    """Per-group threshold grid search to minimize EOD."""
    thresholds = {}
    best_eod = np.inf
    for t in np.linspace(0.01, 0.99, 199):
        fps = []; fns = []
        for g in np.unique(groups):
            mask = groups == g
            pred = probs_all[mask] >= t
            fp = np.mean((pred == 1) & (y_all[mask] == 0))
            fn = np.mean((pred == 0) & (y_all[mask] == 1))
            fps.append(fp); fns.append(fn)
        eod = max(max(abs(fps[g] - fps[ref_group]) for g in range(len(fps))),
                  max(abs(fns[g] - fns[ref_group]) for g in range(len(fns))))
        if eod < best_eod:
            best_eod = eod; thresholds = {g: t for g in np.unique(groups)}
    return thresholds, best_eod

def platt_scale(probs_all, y_all, groups):
    """Group-wise Platt scaling (logistic calibration)."""
    calibrators = {}
    for g in np.unique(groups):
        mask = groups == g
        lr = LogisticRegression().fit(probs_all[mask].reshape(-1, 1), y_all[mask])
        calibrators[g] = lr
    return calibrators

Key Results§

• Significance vs. effect size: Significance ranking is recovered only after normalizing by MDE.
• Mitigation: Threshold optimization is consistently effective; Platt scaling is not — report the distribution of effects.
• Mechanism diagnostic: Silent catastrophic failure under proxy misspecification.
• Drift monitoring: Thresholds are cohort-specific; transfer fails.

Limitations§

All results are synthetic with known ground truth; the pipeline's clinical validity is not established. The benchmark is designed to stress-test components, not to mimic real-world data generation fully.