A machine learning model combining liquid chromatography-mass spectrometry steroid profiling with an interpretable decision-tree algorithm accurately classified congenital disorders of adrenal steroidogenesis, achieving 97% accuracy during internal validation and 99% accuracy in an independent validation cohort, according to a study published in The Journal of Clinical Endocrinology & Metabolism.

"In children—especially newborns and infants—getting the exact subtype diagnosis of congenital disorders of adrenal steroidogenesis (CDAS) quickly can make a real difference in management," corresponding author Tulay Guran, MD, of the Department of Pediatric Endocrinology and Diabetes at Marmara University in Istanbul, Turkey, told AACE Endocrine AI. "These conditions may appear similar at first, but their treatment needs and long-term outcomes can vary considerably. In daily practice, reaching the correct diagnosis may require multiple hormonal tests, stimulation testing, expert interpretation, and genetic confirmation, all of which require time and resources. Our study offers a way to support faster and more accurate etiological classification and help clinicians make earlier, more targeted treatment decisions."

The study included a development cohort of 1,027 participants: 325 patients with genetically confirmed CDAS representing eight disease subtypes and 702 controls. Plasma steroid hormone profiling was performed before treatment initiation using LC-MS/MS. The researchers used the Light Gradient Boosting Machine algorithm to identify the most discriminative steroid hormones and incorporated those features into an optimized machine learning-assisted decision tree.

During internal validation using stratified fivefold cross-validation, the model achieved an overall accuracy of 97%, sensitivity of 99.5%, specificity of 94%, and a macro-average area under the curve of 0.97.

Performance also remained strong across most major CDAS subtypes. For both 17α-hydroxylase/17,20-lyase deficiency and non-congenital adrenal hyperplasia primary adrenal insufficiency, the model achieved 100% accuracy, sensitivity, specificity, and F1-score.

High performance was also observed during internal validation for classical 21α-hydroxylase deficiency (99% accuracy, 99% specificity, and 95% F1-score) and 11β-hydroxylase deficiency (99% accuracy, 100% specificity, and 98% F1-score).

Sensitivity was notably lower for rarer or milder disorders, including aldosterone synthase deficiency (65%), nonclassical 21α-hydroxylase deficiency (71%), 3β-hydroxysteroid dehydrogenase deficiency (72%), and P450 oxidoreductase deficiency (75%).

The researchers evaluated the finalized classifier in an independent validation cohort of 507 additional LC-MS/MS steroid profiles from symptomatic patients undergoing evaluation for premature adrenarche, polycystic ovary syndrome, congenital adrenal hyperplasia, or adrenal insufficiency. Of these, 118 patients had genetically confirmed CDAS, while the remaining patients had nondiagnostic steroid profiles and no confirmed steroidogenic disorder during follow-up.

In the independent cohort, the model achieved 99% overall accuracy, 94% sensitivity, 99.8% specificity, and a 96% F1-score.

The researchers also conducted feature importance and Shapley Additive Explanations (SHAP) analyses to identify which steroid biomarkers contributed most to the model's predictions and to assess model interpretability. The most influential biomarkers included 11-deoxycortisol, 17-hydroxyprogesterone, 21-deoxycortisol, and corticosterone. Elevated 11-deoxycortisol contributed to identifying 11β-hydroxylase deficiency, whereas elevated 17-hydroxyprogesterone and 21-deoxycortisol strongly supported predictions of classical 21α-hydroxylase deficiency.

Visualization analyses demonstrated distinct biochemical clustering among most CDAS subtypes. Controls formed a separate compact cluster from affected patients, supporting that the model captured biologically meaningful steroid patterns rather than statistical artifacts. Exploratory analyses found no significant differences in model performance across age groups or between sexes.

The researchers acknowledged several limitations. Some rare CDAS subgroups contained relatively few patients, limiting the precision of subtype-specific performance estimates. Several steroid variables required imputation because of incomplete analytical availability, introducing potential uncertainty. In addition, the model was developed using data from a single center.

The researchers reported no conflicts of interest.

AACE Endocrine AI invited Dr. Guran, MD to elaborate on the study's findings.

What data stood out? 

Dr. Guran: We already knew from clinical experience that LC-MS/MS steroid profiling has strong diagnostic power when interpreted by an expert because these disorders have distinct biochemical signatures. We also expected that a machine learning model could recognize these patterns even without incorporating clinical findings. However, what was particularly encouraging was the high sensitivity and overall performance the model achieved, even beyond what we had anticipated. It was reassuring to see that a single steroid profile could distinguish several congenital adrenal steroidogenesis disorders with strong accuracy, not only in the development cohort but also in an independent validation cohort. This gave us confidence that the model was capturing true pathophysiological signatures rather than nonspecific statistical patterns.

How might the findings influence endocrinology clinical practice?

Dr. Guran: In clinical practice, LC-MS/MS steroid profiling is a powerful diagnostic tool, but interpreting multisteroid patterns can be challenging and often depends on specialized expertise. A transparent machine learning-assisted decision tree could make this interpretation faster, more standardized, and more accessible for clinicians. This may be particularly useful in newborns, infants, or patients with atypical or overlapping biochemical findings, in whom early diagnostic orientation is critical. By helping clinicians recognize the most likely etiological subtype earlier, the model could shorten the diagnostic pathway, guide the selection of confirmatory tests, and support timely, targeted treatment decisions. At the same time, we see this approach as a decision-support tool that complements—not replaces—clinical assessment, ACTH stimulation testing when indicated, and genetic confirmation.

Is there anything else you'd like to say about this work?

Dr. Guran: We see this work as an important step toward precision endocrinology. Congenital disorders of adrenal steroidogenesis are rare and heterogeneous, and even experienced clinicians may face diagnostic challenges when biochemical findings overlap or when the clinical presentation is atypical. By combining steroid signatures with interpretable AI, we aimed to develop a tool that is both accurate and clinically understandable. Of course, further validation in larger, multicenter, and international cohorts will be essential. Our hope is that, in the future, approaches like this can be integrated into endocrine laboratory reports or clinical decision-support systems to help clinicians reach timely diagnoses and improve patient care.