BACKGROUND: Early and accurate etiological diagnosis of congenital disorders of adrenal steroidogenesis (CDAS) is critical as timely targeted management can prevent life-threatening complications and improve long-term outcomes. OBJECTIVE: To develop and validate a machine learning-assisted decision tree model for classifying CDAS using plasma steroid hormone profiles quantified by liquid chromatography-mass spectrometry (LC-MS/MS). METHODS: A development cohort of 1027 participants (325 genetically confirmed CDAS patients representing eight subtypes/702 controls) was used for model construction. The Light Gradient Boosting Machine algorithm identified key discriminatory steroid hormones, which were integrated into an optimized decision-tree classifier. Internal performance was assessed through five-fold cross-validation. The performance of the model was further evaluated using a validation cohort comprising 507 independent LC-MS/MS steroid profiles. Additional analyses included Shapley additive explanations (SHAP), confusion matrix visualization, Principal Component Analysis (PCA), and Uniform Manifold Approximation and Projection (UMAP). RESULTS: In the development cohort, the model achieved a mean overall accuracy of 97.1%, sensitivity of 99.5%, and specificity of 93.7%, with a macro-AUC (area under curve) of 0.97. Subtype-level accuracy exceeded 98% for most major CDAS subtypes. In the validation cohort, overall accuracy was 98.9%, sensitivity 93.6%, specificity 99.8%. Feature importance analysis and SHAP identified 11-deoxycortisol, 17-hydroxyprogesterone, 21-deoxycortisol, and corticosterone as the strongest discriminators. PCA and UMAP revealed distinct clustering of CDAS subtypes, confirming the biological coherence of model predictions. CONCLUSION: Machine learning-assisted steroid profiling provides an accurate and highly interpretable diagnostic approach for CDAS, with potential for integration into pediatric endocrine diagnostics and decision-support systems.
Tosun et al. (Thu,) studied this question.