Early detection of G2SCH through machine learning analysis of physical examination metrics

Grade 2 subclinical hypothyroidism (G2SCH) is associated with an increased risk of various diseases but is rarely detected before symptom onset. This study aims to develop a machine learning-based prediction model for G2SCH using routine physical examination data. 57,539 samples were collected from one institution for training and 33,880 from another one for testing. Three feature-selecting methods were employed, while five machine learning algorithms were utilized. SHapley Additive exPlanation (SHAP) index of selected features were calculated. Area under curve (AUC), sensitivity and specificity were used to evaluate models. Our results demonstrate that the logistic regression model, using features selected by LASSO under a 65:35 SMOTENC ratio, achieved robust predictive performance with an AUC of 0.870 (95% CI: 0.834, 0.906), sensitivity of 86.8%, and specificity of 70.8%. We developed a high-performance machine learning model for G2SCH prediction based on physical examination data, enabling timely referral and appropriate treatment of G2SCH patients.