ESPE Abstracts

Objective: The models of diagnosis of central precocious puberty (CPP) in girls were constructed based on machine learning algorithm to assist the management of CPP diagnosis in girls.

Methods: Girls who visited the department of endocrinology, genetics and metabolism of the Fuzhou Children's Hospital of Fujian Medical University from January 2014 to June 2023 and were consistent with the diagnosis of precocious puberty (PP) were retrospectively included. They had completed the gonadotropin-releasing hormone stimulation test. The included children were randomly divided into three data sets: training set (80%), validation set (10%), and test set (10%). The training set data was used to screen the predictors based on Lasso regression. Validation set data assisted the training set data to adjust the model parameters to establish the girl CPP diagnosis prediction model based on five machine learning algorithms including Logistic regression, support vector machine, random forest, limit gradient lifting and artificial neural network, and carried out internal validation. The test set data was used for model external validation.

Results: A total of 1310 girls with PP were analyzed, of which 502 (38.3%) were diagnosed with CPP and 808 (61.7%) were excluded from CPP diagnosis. Lasso regression screened out eight predictors, including: age (y), group of disease course (course<0.5 y, 0.5 y≤course<1.0 y, course≥1.0 y), group of basal luteinising hormone (LH) (LH<0.20 IU/L, 0.20 IU/L≤LH<0.83 IU/L, LH≥0.83 IU/L), bone age (BA) ahead (y), height SDS for BA, uterus volume (ml), and larger ovary volume (ml). Five machine learning algorithm prediction models were developed. The support vector machine model performed best in both internal and external validation. The areas under the receiver operating characteristic (ROC) curve were 0.829 and 0.873, respectively, and the accuracy rates were 78.6% and 80.9%, respectively. When the prediction probability is greater than 0.75 as the CPP diagnostic cut-off point, the specificity of the model in the validation set and the test set data were 96.4% and 97.6%.

Conclusion: This study developed prediction models based on machine learning algorithm for the diagnosis of CPP girls. Internal and external validation ensured a good degree of discrimination and calibration of the models, which could assist the diagnosis of CPP under the condition of basic clinical information.