Identification of Metabolic Phenotypes in Polycystic Ovary Syndrome Using Unsupervised Clustering: A Case-Control Study

Abstract

PCOS is an endocrine disorder of women. It remains unclear whether adding biomarkers to clinical assessment improves diagnosis, or if androgens explain clinical hirsutism severity. Objectives: To identify metabolic phenotypes via clustering; test if a biomarker panel predicts PCOS better than clinical assessment; and explore the serum androgen-hirsutism correlation. Methods: This retrospective case-control study analyzed Kaggle data (208 PCOS cases and controls). Clinical, biochemical (total/free testosterone, DHEAS, LH/FSH, HOMA-IR, SHBG), and ultrasound (ovarian volume, follicle count) variables were recorded. Logistic regression (clinical vs. full panel) was compared using AUC, DeLong's test, LRT, bootstrap correction, and Hosmer-Lemeshow. K-means clustering (k=2) applied to standardized BMI, HOMA-IR, and SHBG. Pearson correlation assessed FG score. Results: The full model (test AUC 0.596) did not outperform the clinical model (AUC 0.592; DeLong p=0.886; LRT p=0.835). Optimism correction reduced the full AUC to 0.507 versus 0.526 for clinical, with poor calibration (p=0.005). No biomarker independently predicted PCOS. No serum hormone correlated with FG score (|r|≤0.111). Clustering revealed obese/low-SHBG (n=98) and lean/insulin-resistant (n=110) phenotypes, differing in BMI, SHBG, and HOMA-IR (p<0.050), but not testosterone or LH/FSH. Conclusions: Adding biomarkers does not improve diagnosis and is overfitted. Serum androgens do not explain hirsutism. Clustering identifies distinct metabolic subtypes, supporting phenotype-based management.