Development of Prediction Models Using Machine Learning Algorithms for Girls with Suspected Central Precocious...

Area of Science:

• Pediatric Endocrinology
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Central precocious puberty (CPP) negatively impacts girls' development.
• Current diagnostic tests, like the GnRH-analogue (GnRHa) test, are costly and invasive.

Purpose of the Study:

• To develop machine learning models for predicting GnRHa test response in girls with CPP.
• To identify key clinical and laboratory features associated with CPP diagnosis.

Main Methods:

• Retrospective analysis of 1757 girls' data undergoing GnRHa testing.
• Development of XGBoost and random forest classifiers.
• Utilized Local Interpretable Model-Agnostic Explanations (LIME) for model interpretability.

Main Results:

• XGBoost and random forest models demonstrated high predictive performance (AUC 0.88-0.90).
• Key predictors identified: basal luteinizing hormone, follicle-stimulating hormone, and insulin-like growth factor-I.
• LIME analysis confirmed the significant contribution of these factors.

Conclusions:

• Developed prediction models can aid in CPP diagnosis.
• These models show potential as a prescreening tool before GnRHa stimulation testing.