Development of Clinical Prediction Models for Girls with Central Precocious Puberty in China: Machine learning Approaches and Interpreting with Explainable Artificial Intelligence

Xiaoxiao Zhu¹Weirong Xue¹Guangshuang Zhou¹Jiajia Qin¹Meng Zhao²Bin Wu³Shengli Li²Yingliang Jin^1*

• ¹Department of Biostatistics, School of Public Health, Xuzhou Medical University, Xuzhou, Jiangsu, China., Xuzhou, Jiangsu, China., China
• ²The Affiliated Hospital of Xuzhou Medical University, Xuzhou, Jiangsu Province, China
• ³Anhui Provincial Children’s Hospital, Hefei, Anhui Province, China

Background and objective: 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. Our objective is to develop a machine learning (ML) model to predict the response of girls suspected of CPP to stimulation tests before injection of GnRHa. Methods: We conducted a retrospective review of 320 girls suspected of CPP who visited the Affiliated Hospital of Xuzhou Medical University and Anhui Provincial Children’s Hospital from 2019 to 2023. We utilized 11 variables (including essential clinical data, baseline hormone data, and pituitary length) to construct the first basic clinical model. Additionally, we developed a second model by incorporating 6 additional variables extracted from pelvic ultrasound reports into the first basic clinical model, aiming to evaluate the incremental value of pelvic ultrasound in the prediction of CPP. Furthermore, we utilized 3 explainable artificial intelligence (XAI) methods to elucidate the optimal model. Results: Among 320 girls suspected of CPP, 188 (58.75%) were diagnosed with CPP, while 132 (41.25%) were diagnosed with non-CPP. Across 9 ML models, the multilayer perceptron (MLP) emerged as the optimal model, delivering outstanding performance. Moreover, pelvic ultrasound demonstrated incremental value in predicting CPP, with uterine length contributing more significantly to prediction model than uterine volume. Consistent findings from Shapley additive explanations (SHAP), local interpretable model-agnostic explanations (LIME), and Break Down methods revealed that difference between bone age and chronological age (BA-CA), chronological age (CA), and basal luteinizing hormone (LH) were the three most important features. Conclusion: Our study is the first to integrate LIME, SHAP, and Break Down methods for interpreting the CPP prediction model. The MLP exhibited overall good performance, serving as a pre-screening tool to reduce unnecessary procedures and facilitate clinicians in making more informed decisions. LIME, SHAP, and Break Down demonstrated robust performance as XAI methods.