Development and validation of diagnostic prediction models for central precocious puberty in girls based on machine learning: a multicenter retrospective study.

Objective: We developed novel diagnostic prediction models for central precocious puberty (CPP) diagnosis in girls based on machine learning. Methods: In this retrospective study, 2148 girls with PP from 4 centers in China were included, and divided into training group (n = 1048), validation group (n = 262), test group 1 (n = 270), test group 2 (n = 278), and test group 3 (n = 290). Diagnostic prediction models were developed based on logistic regression, linear support vector machine (SVM), RandomForest, XGBoost and fully connected neural network. Model discrimination was assessed with C-statistics and accuracy rates. Model calibration was analyzed by calibration curves. Results: A total of eight independent predictors were screened by Lasso regression, including chronological age, disease course, height SDS, basal luteinizing hormone, bone age (BA) - chronological age (CA), height SDS for bone age, uterine volume, and larger ovarian volume. Among the five established prediction models, the SVM model demonstrated optimal performance, achieving AUC values of 0.850 in internal validation and 0.827 in external validation, accuracy rates of 78.6% in internal validation and 72.1% average in external validations. The model is displayed on the website: https://wuwenyong.shinyapps.io/CPPpredict/. Conclusions: The SVM prediction model can assist diagnose CPP. Through the internal and multi-center external validations, the model showed good degree of discrimination and calibration.