Early phototherapy prediction for neonatal hyperbilirubinemia: a machine and deep learning-based approach for Indian cohorts

Neonatal hyperbilirubinemia is a common condition that can lead to severe neurological complications if untreated. Currently used nomograms for predicting phototherapy needs are based on few cohorts and may not generalize across populations, particularly Indian neonates. The study was aimed to develop and evaluate machine and deep learning models to predict the need for phototherapy using readily available clinical and demographic data. Data from 545 Indian neonate–mother dyads were collected from January to August 2024. Variables included neonatal and maternal risk factors, bilirubin levels, gestational weeks, and cord thyroid-stimulating hormone (TSH). After data preprocessing and class balancing using SMOTE, seven machine learning and three deep learning models were trained and evaluated. The Random Forest Classifier demonstrated the highest predictive accuracy (83%) with an AUC of 0.91, followed by the Gradient Boost Classifier (82%), TabNet Classifier (81%), and Liquid Neural Network (80%). SHAP-based interpretability analyses identified bilirubin levels and cord TSH as the most influential features. The findings highlight the potential of deep learning for early phototherapy prediction tailored to Indian populations; this approach can support timely interventions in neonatal care.