Machine learning identifies girls with central precocious puberty based on multisource data.

Area of Science:

• Pediatric Endocrinology
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Central precocious puberty (CPP) diagnosis typically relies on the gonadotropin-releasing hormone (GnRH) stimulation test, which is invasive and resource-intensive.
• There is a need for simplified, non-invasive diagnostic tools to identify girls at high risk of CPP earlier and more efficiently.

Purpose of the Study:

• To develop and validate simplified diagnostic models for identifying central precocious puberty (CPP) in girls.
• To evaluate the efficacy of machine learning models using readily available clinical data, avoiding the need for the standard GnRH stimulation test.

Main Methods:

• Female patients with early onset of secondary sexual characteristics underwent data collection, including clinical visits, laboratory tests, and medical imaging.
• Features were extracted from unstructured data (reports, images), and Extreme Gradient Boosting (XGBoost) models were trained using single-source and multisource data.
• Classification of patients into CPP or non-CPP groups was performed using these machine learning models.

Main Results:

• A multisource data model achieved an area under the curve (AUC) of 0.88 and a Youden index of 0.64.
• Analysis of single-source models indicated that basal hormone tests provided the highest diagnostic value for CPP.
• Feature importance analysis highlighted key variables contributing to accurate CPP diagnosis.

Conclusions:

• Three simplified diagnostic models were successfully developed using easily accessible clinical data prior to the GnRH stimulation test.
• These models can effectively identify girls at high risk of CPP, offering tailored solutions for various clinical settings.
• The integration of machine learning and multisource data fusion presents a promising advancement over traditional diagnostic methods for CPP.