Objective
The current diagnosis of central precocious puberty (CPP) relies on the gonadotropin-releasing hormone analogue (GnRHa) stimulation test, which requires multiple invasive blood sampling procedures. The aim of this study was to construct machine learning models incorporating basal pubertal hormone levels, pituitary magnetic resonance imaging (MRI), and pelvic ultrasound parameters to predict the response of precocious girls to GnRHa stimulation test.
Methods
This retrospective study included 455 girls diagnosed with precocious puberty who underwent transabdominal pelvic ultrasound, brain MRI examinations and GnRHa stimulation testing were retrospectively reviewed. They were randomly assigned to the training or internal validation set in an 8:2 ratio. Four machine learning classifiers were developed to identify girls with CPP, including logistic regression, random forest, light gradient boosting (LightGBM), and eXtreme gradient boosting (XGBoost). The accuracy, sensitivity, specificity, positive predictive value, negative predictive value, area under receiver operating characteristic (AUC) and F1 score of the models were measured.
Results
The participates were divided into an idiopathic CPP group (n = 263) and a non-CPP group (n = 192). All machine learning classifiers used achieved good performance in distinguishing CPP group and non-CPP group, with the area under the curve (AUC) ranging from 0.72 to 0.81 in validation set. XGBoost had the highest diagnostic efficacy, with sensitivity of 0.81, specificity of 0.72, and F1 score of 0.80. Basal pubertal hormone levels (including luteinizing hormone, follicle-stimulating hormone, and estradiol), averaged ovarian volume, and several uterine parameters were predictors in the model.
Conclusion
The machine learning prediction model we developed has good efficacy for predicting response to GnRHa stimulation tests which could help in the diagnosis of CPP.