Interpretable machine learning for the non-invasive differentiation of central precocious puberty from premature thelarche in girls: a diagnostic modeling study to reduce reliance on the gonadotropin-releasing hormone stimulation test
Background The gonadotropin-releasing hormone (GnRH) stimulation test remains the reference standard for confirming central precocious puberty (CPP), but is invasive and costly. We developed and validated an interpretable machine-learning model differentiating CPP from premature thelarche (PT) using only pre-stimulation information. Methods From 15,385 consecutive outpatients (2020–2026) at a tertiary hospital in northwest China, we analyzed 130 girls with early breast development, a final diagnosis of CPP (n=51) or PT (n=79), and basal luteinizing hormone (LH) and follicle-stimulating hormone (FSH) measurements. Seventeen candidate pre-stimulation predictors were reduced by LASSO regression. Six algorithms were compared by repeated stratified 5-fold cross-validation, with feature selection, imputation, and standardization refitted inside each training fold (fully nested validation). The final model was evaluated by out-of-fold AUC, calibration, decision-curve analysis, SHAP, and a pre-specified two-threshold triage. Results The most frequently selected predictors were age, basal FSH, and the basal LH/FSH ratio. Penalized logistic regression achieved a nested AUC of 0.813 (pooled out-of-fold 0.805, 95% CI 0.729–0.876), indistinguishable from non-linear comparators. Age alone achieved 0.818; adding hormonal variables raised the AUC to 0.855 and, at matched 90% sensitivity, specificity from 0.569 to 0.627. Calibration was moderate (slope 0.600). Under the triage rule, 38.5% of girls fell below the rule-out threshold (negative predictive value 0.900; 5 of 50 with CPP) and 21.5% exceeded the rule-in threshold (positive predictive value 0.750); an estimated 60.0% could potentially have deferred the stimulation test—a retrospective estimate, not a demonstrated reduction. Discrimination was attenuated at ages 6–8.99 years (AUC 0.707) and stable in a hardened reference-standard subset (n=101; AUC 0.787). Conclusions A parsimonious, hand-computable three-variable model using only pre-stimulation data stratifies girls with early breast development and could potentially defer the GnRH stimulation test in an estimated three-fifths, at the cost of initially assigning 5 girls with CPP to structured surveillance. External multicenter prospective validation with local recalibration is required before clinical use.
Authors
- Nan Nan (ORCID: https://orcid.org/0000-0002-6855-8268)
- Lei Huang (ORCID: https://orcid.org/0000-0002-9279-878X)
- Ying Zhu
- Zhenqiang Da
- Xingwen Xie
- Yang Gao
- Jiaxin Chai
- Weikai Wang
- Yanan Liu
- Yaxiong Gao
- Jiayue Li
- Tingting Li
- Dandan Lu
Institutions
- Gansu University of Traditional Chinese Medicine (CN)
- Gansu Provincial Maternal and Child Health Hospital (CN)
- Gansu Institute of Political Science and Law (CN)
- Gansu Provincial Hospital of TCM (CN)
- Lanzhou Institute of Technology (CN)
Publication Details
- Journal
- Frontiers in Endocrinology
- Published
- 2026-11-16
- DOI
- https://doi.org/10.3389/fendo.2026.2008437
- Primary Topic
- Hypothalamic control of reproductive hormones
- Type
- article
- Field-Weighted Citation Impact
- 0.00