Development of Clinical Diagnostic Models for Girls with Central Precocious Puberty in China: Machine-learning Approaches and Interpreting with XAI (Preprint)

BACKGROUND Background: The incidence of central precocious puberty (CPP) in girls has been steadily increasing. However, stimulation tests remain costly, time-consuming, and painful. Despite efforts to simplify these tests or identify alternative biomarkers, consensus has not been achieved. OBJECTIVE Our objective was to develop a machine learning (ML) assisted diagnostic model for CPP, predicting girls' responses to stimulation tests to avoid unnecessary procedures. METHODS Methods: We conducted a retrospective review of 320 girls treated at Xuzhou Medical University Affiliated Hospital and Anhui Provincial Children's Hospital from 2019 to 2023, who underwent GnRH agonist stimulation tests. We utilized 11 variables including clinical data, bone age (BA), pituitary length, and incorporated six variables extracted from pelvic ultrasound reports to assess the value of pelvic ultrasound in CPP diagnosis. The optimal model was selected based on comprehensive metrics from training and validation data. Additionally, we employed three descriptive ML interpretability methods to elucidate the model. RESULTS Results: Among 320 girls suspected of CPP,188(58.75%) were diagnosed with CPP. The multi-layer perceptron (MLP) model demonstrated excellent performance metrics, sensitivity (0.969), specificity (1), positive predictive value (PPV) (1), negative predictive value (NPV) (0.958), F1-score (0.9846), Matthews correlation coefficient (MCC) (0.964), accuracy (0.982), and achieved a high area under the receiver operating characteristic curve (AUC) of 0.9908, and area under the precision-recall curve (AUPRC) of 0.9946. Pelvic ultrasound showed supportive roles in CPP diagnosis, with uterine length contributing more significantly than uterine volume. Consistent results from SHAP, LIME, and Break Down methods indicated that BA, chronological age (CA), and basal luteinizing hormone (LH) were the three most important factors. CONCLUSIONS Conclusion: Our study is the first to integrate LIME, SHAP, and Break Down methods for CPP diagnosis. The predictive model we developed exhibited overall good performance, serving as a pre-screening tool before stimulation tests to reduce unnecessary procedures and aid clinicians in making informed decisions. LIME, SHAP, and DALEX demonstrated robust performance as explainable artificial intelligence (XAI) methods.