Artificial Intelligence Neonatal Jaundice Risk Prediction Under Variable Transcutaneous Bilirubinometer Skin-Tone Calibration: A Fairness-Audit Evaluation of Healthcare Phototherapy Initiation

Transcutaneous bilirubinometry is the front-line screening instrument for neonatal hyperbilirubinaemia, and the calibration curves that convert its readings into serum-equivalent bilirubin concentrations are known to vary with skin tone. When an artificial intelligence risk model is layered on top of these readings to recommend phototherapy initiation, the calibration variability propagates into the model's decision boundary in ways that a standard discrimination audit cannot detect. The framework treats skin-tone-dependent calibration as a structured perturbation of the input distribution and defines a calibration-conditional risk gap as the difference in expected decision loss between calibration strata at a fixed clinical threshold. A variance-decomposition argument shows that the gap is identifiable from routinely collected audit data only when the calibration stratum is observed, and that unstratified audits systematically understate it. The design parameters used to characterise the audit's power are a between-cluster standard deviation of 6.0, a residual standard deviation of 4.5, an intraclass correlation of 0.64, and cluster counts of 8, 16, 24 and 32 with effect sizes of 3, 5 and 8 percentage points across 2000 replications per cell. The analysis indicates that 56 clusters are required to detect a 5-percentage-point calibration-conditional gap at 80% power. The contribution is a formal statement of what a fairness audit of phototherapy-initiation models can and cannot establish under realistic calibration heterogeneity.