Clinical Computer Vision for Neonatal Jaundice Screening Under Variable Skin Tone Calibration: A Fairness-Audit Evaluation of Transcutaneous Bilirubin Threshold Errors Across District Hospitals

Clinical computer vision systems that estimate transcutaneous bilirubin from smartphone photographs promise to extend neonatal jaundice screening to district hospitals where laboratory bilirubinometry is unavailable, but their threshold performance is calibrated on device-specific colour references that do not transfer across skin tones. The framework separates calibration error into a device-referenced component and a skin-tone-referenced component, defines threshold error as a function of both, and derives the conditions under which a pooled audit estimate masks subgroup-specific miscalibration. The design evaluation used a before-after-control-impact estimator with between-cluster SD = 6.0, residual SD = 4.5 and intraclass correlation = 0.64, across cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points, with 2000 replications per cell. Power reached 80% at 52 clusters for a 5-percentage-point effect. The framework proposes that fairness audits of bilirubin threshold error must be stratified by skin tone before pooling, and that the before-after-control-impact design is adequate only when cluster counts exceed the derived threshold.