Using machine learning to discriminate non-classical 21-hydroxylase deficiency from polycystic ovary syndrome: an...

Area of Science:

• Endocrinology
• Biochemistry
• Machine Learning in Medicine

Background:

• Nonclassical 21-hydroxylase deficiency (NC21OHD) is often misdiagnosed as polycystic ovary syndrome (PCOS) due to similar symptoms.
• Current diagnostic methods like the cosyntropin stimulation test are invasive and impractical for routine use.
• Previous research developed a machine learning model using basal serum steroid profiles to identify NC21OHD non-invasively.

Purpose of the Study:

• To validate and refine a machine learning model for distinguishing NC21OHD from PCOS.
• To evaluate the model's diagnostic performance using baseline steroid signatures.
• To establish a simpler, non-invasive diagnostic approach for NC21OHD.

Main Methods:

• A tricentric study involving 447 women (PCOS and NC21OHD) for training and validation.
• Quantification of 20 serum steroids using liquid chromatography-tandem mass spectrometry (LC-MS/MS).
• Analysis via orthogonal partial least squares discriminant analysis (OPLS-DA) with internal and external validation.

Main Results:

• The 20-steroid machine learning model achieved 100% accuracy in separating NC21OHD from PCOS in validation cohorts.
• Key discriminant metabolites identified were 21-deoxycortisone, 21-deoxycortisol, and 17-hydroxyprogesterone.
• Simplified 3- and 6-steroid models showed slightly reduced sensitivity (∼90%) and specificity (∼97%).

Conclusions:

• Machine learning combined with basal LC-MS/MS steroid profiling accurately identifies NC21OHD without requiring dynamic stimulation tests.
• This non-invasive approach can simplify diagnosis, enhance patient comfort, and facilitate large-scale screening.
• The model offers a promising alternative for diagnosing hyperandrogenic conditions.