Newborn Screening for Thyroid Dysfunction Through Explainable SHAP Feature Selection and Stacking Ensemble With LLM Assisted Diagnosis
Ying Zhou, Jiawen Zhang, Chi Chen, Huaqing Mao, Rulai Yang · IEEE Access · 2025
Newborn screening (NBS) for thyroid dysfunction traditionally relies on physician interpretation of laboratory results, which can be subjective and labor-intensive. In this study, a two-phase machine learning framework was developed to automate the diagnostic process. Based on screening data from 5,406 newborns collected from Children’s Hospital ZheJiang University School of Medicine, 7 thyroid-related indicators were analyzed. In the first phase, a Random Forest model was trained to distinguish healthy individuals from those with thyroid dysfunction, achieving an AUC of 0.9914, a sensitivity of 98.57%, and a specificity of 98.60%. In the second phase, diseased samples were further classified into congenital hypothyroidism (CH), elevated TSH, or elevated TSH requiring medication, with a stacking ensemble model achieving a macro AUC of 0.9696 and an accuracy of 93.75%. SHAP value analysiswas used both for feature selection and interpretability, identifying TSH, FT4, and T4 as the most informative biomarkers. A large language model (LLM) was integrated to assist cases with low model confidence. The system was deployed through a web-based interface, enabling automatic diagnosis based on input laboratory indicators.