Neonatal jaundice (hyperbilirubinemia) affects 60%–80% of term and over 80% of preterm neonates worldwide; unmanaged pathological cases risk kernicterus and irreversible neurological damage. Traditional detection relies on invasive total serum bilirubin (TSB) measurement or subjective visual assessment, both poorly scalable in resource-limited settings. This PRISMA-compliant review synthesises 65 peer-reviewed computational studies (January 2010–March 2026), spanning 18 traditional machine learning, 17 deep learning, 13 smartphone/non-invasive system, 9 electronic health record (EHR)-based and 8 multimodal studies. Methods reviewed include SVM, Random Forest, XGBoost, CNNs, Vision Transformers, Graph Neural Networks and capsule networks. Key accuracy findings include: the best image-based model (T2T-ViT) achieved AUC = 0.99 and 99% accuracy; a Capsule Graph Neural Network optimised with the Wild Horse Algorithm reached 98.52% accuracy (Nelson et al., 2024); the only prospective multi-ethnic validation (Boo et al., 2024) demonstrated 100% sensitivity and AUC = 0.89 across three ethnic groups; and the largest EHR-based model (Chou et al., 2020a) achieved a mean absolute error of 1.05 mg/dL on 27,428 records. A novel deployment-readiness scoring framework reveals that only 3 of 65 studies meet clinical deployment thresholds, with skin phototype coverage (Fitzpatrick Types V–VI) and prospective multi-site validation as the most frequently unmet criteria. Challenges including skin phototype bias, illumination variability, model interpretability and absence of regulatory (FDA/CE/CDSCO) approval pathways are systematically analysed. Future directions include federated learning, self-supervised and few-shot paradigms, hyperspectral imaging for melanin-independent bilirubin sensing and vision-language foundation models. These are identified as critical enablers for bridging the gap between laboratory proof-of-concept and clinically deployable neonatal jaundice screening.
Achaliya et al. (Mon,) studied this question.