Clinical Computer Vision for Neonatal Jaundice Severity Estimation Under Variable Skin Tone Calibration: A Fairness-Audit Evaluation of Transcutaneous Bilirubin Threshold Safety

Clinical computer vision systems that estimate neonatal jaundice severity from skin images must operate across a wide range of skin tones, yet the safety of the transcutaneous bilirubin thresholds they inform has not been characterised under realistic calibration error. The framework is formalised as a threshold-safety model in which calibration bias shifts the effective decision boundary, and the resulting false-negative and false-positive rates are expressed as functions of bias magnitude, threshold margin, and the population distribution of bilirubin values. A replication-based evaluation of the before-after-control-impact estimator was conducted to characterise the design's operating characteristics under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), with power estimated across 2000 design cells. A calibration bias of 2 mg/dL can eliminate the safety margin of a 15 mg/dL referral threshold for a substantial fraction of the neonatal population, and that the bias required to produce a given false-negative inflation depends jointly on the threshold margin and the local prevalence of severe hyperbilirubinaemia. The framework provides a structured method for auditing threshold safety before clinical deployment and identifies the conditions under which skin-tone calibration error becomes clinically consequential.