Identification and Optimization of Contributing Factors for Precocious Puberty by Machine/Deep Learning Methods in Chinese Girls

Background And Objectives: As the worldwide secular trends are toward earlier puberty, identification of contributing factors for precocious puberty is critical. We aimed to identify and optimize contributing factors responsible for onset of precocious puberty machine learning/deep learning algorithms in girls.

Methods: A cross-sectional study was performed among girls aged 6-16 years from 26 schools in Beijing based on a cluster sampling method. Information was gleaned online questionnaires. Machine/deep learning algorithms were performed using Python language (v3.7.6) on PyCharm platform.

Results: Of 11308 students enrolled, there are 5527 girls, and 408 of them had experienced precocious puberty. Training 13 machine learning algorithms revealed that gradient boosting machine (GBM) performed best in predicting precocious puberty. By comparison, six top factors including maternal age at menarche, paternal body mass index (BMI), waist-to-height ratio, maternal BMI, screen time, and physical activity were sufficient in prediction performance, with accuracy of 0.9530, precision of 0.9818, and area under the receiver operating characteristic curve (AUROC) of 0.7861. The performance of the top six factors was further validated by deep learning sequential model, with accuracy reaching 92.9%.

Conclusions: We identified six important factors from both parents and girls that can help predict the onset of precocious puberty among Chinese girls.