Related Experiment Video
Updated: Jul 9, 2026
Identification of Disease-related Spatial Covariance Patterns using Neuroimaging Data
Published on: June 26, 2013
Using machine learning to discriminate non-classical 21-hydroxylase deficiency from polycystic ovary syndrome: an
T Lecot-Connan1,2,3, G Bachelot1, B Donadille4
1Département de Métabolomique Clinique, Hôpital Saint Antoine, AP-HP. Sorbonne Université, 27 Rue Chaligny, Paris 75012, France.
Introduction:
Nonclassical 21-hydroxylase deficiency (NC21OHD) is a rare autosomal recessive disorder that is frequently misdiagnosed as polycystic ovary syndrome (PCOS) because of their overlapping clinical and biochemical presentations. The conventional diagnostic method, the cosyntropin stimulation test, is invasive, time consuming, and difficult to implement in routine clinical practice. To overcome these limitations, we previously developed a proof-of-concept machine learning model integrating basal serum steroid profiling by liquid chromatography with tandem mass spectrometry (LC-MS/MS) to identify NC21OHD without the need for dynamic stimulation testing. The present study aimed to validate and refine this model in a larger, multicentric cohort to evaluate its diagnostic performance in distinguishing NC21OHD from PCOS based on baseline steroid signatures.
Methods:
This tricentric study included 447 women: 263 with PCOS and 59 with NC21OHD in the training set (Pitié-Salpêtrière Hospital, Paris), and 104 PCOS and 21 NC21OHD in 2 independent validation cohorts (Saint-Antoine Hospital, Paris; Bordeaux University Hospital). Twenty serum steroids were quantified by LC-MS/MS and analyzed using orthogonal partial least squares discriminant analysis. Model robustness was assessed by internal cross-validation and external validation.
Results:
20-steroid model achieved complete separation between NC21OHD and PCOS in both external validation cohorts (accuracy = 100%). The most discriminant metabolites were 21-deoxycortisone, 21-deoxycortisol, and 17-hydroxyprogesterone. Simplified 3- and 6-steroid models demonstrated reduced sensitivity (∼90%) and specificity (∼97%).
Conclusion:
Machine learning combined with basal LC-MS/MS steroid profiling enables accurate identification of NC21OHD without dynamic stimulation. This approach could simplify diagnosis, improve patient comfort, and support large-scale screening of hyperandrogenic women.
