Neonatal jaundice is a frequent clinical condition that requires timely monitoring to reduce the risk of severe hyperbilirubinemia-related complications. Although serum bilirubin measurement remains the clinical reference standard, image-based screening approaches may support non-invasive and scalable monitoring workflows. This study presents a chromatic feature-based computer vision workflow for neonatal jaundice screening from digital images. A pre-existing image dataset of neonates with and without jaundice was used to develop and evaluate a retrospective computational proof-of-concept. The proposed workflow integrates automatic skin segmentation, region-of-interest extraction, chromatic feature computation from the YCbCr and CIE Lab color spaces, feature selection through an ablation study, and binary classification using XGBoost. A probability calibration stage and decision-threshold adjustment were incorporated to prioritize sensitivity, considering the screening-oriented nature of the task. Model performance was assessed using stratified five-fold cross-validation. The workflow achieved an AUC-ROC of 0.867, a sensitivity of 84.7%, a specificity of 69.4%, and a negative predictive value of 93.9%. These results suggest that combined YCbCr-CIE Lab chromatic descriptors provide useful discriminative information for image-based neonatal jaundice screening. However, the study should be interpreted as a computational proof of concept rather than a clinically validated diagnostic system. External validation using prospectively acquired images under diverse acquisition conditions is required before translation into clinical or remote monitoring environments.

Keywords

Computer vision; Neonatal jaundice; Chromatic feature analysis