Development and comparison of machine learning-based clinical diagnostic models for girls with suspected central precocious puberty

Central precocious puberty (CPP) diagnosis requires invasive GnRH stimulation testing. We developed and compared five machine learning models using routinely available clinical, laboratory, and ultrasound data from 1,033 girls (237 CPP). After systematic feature selection, 20 features were retained from 24 candidates. Models were evaluated through 20 repeated stratified evaluations with comprehensive assessment of discrimination, calibration, and clinical utility. CatBoost achieved the highest discrimination (AUROC: 0.894 ± 0.024) with satisfactory calibration (ECE: 0.059 ± 0.010), with competitive discrimination against traditional diagnostic approaches and notably better calibration than deep learning baselines. Random serum luteinizing hormone was the dominant predictor (mean |SHAP| = 1.267). At the optimal threshold, the model could avoid 63 unnecessary GnRH tests per 100 cases while maintaining 84% sensitivity with a negative predictive value of 94.8%. These results support CatBoost as a potential pre-screening tool to reduce unnecessary invasive testing in girls with suspected CPP.