Unsupervised subtyping of polycystic ovary syndrome: current validity and priorities for clinical translation

Polycystic ovary syndrome (PCOS), recently renamed polyendocrine metabolic ovarian syndrome (PMOS), encompasses substantial androgenic, reproductive, metabolic and life-course heterogeneity. Unsupervised analyses of patient data have identified related metabolic and reproductive or neuroendocrine profiles, although their composition and stability vary across studies. In this Review, we compare Rotterdam A-D phenotypes with the Dapas model, its application to Rotterdam-defined cases by van der Ham, and the Gao system; examine study-level choices in phenotyping, hormone assays, age and body mass index handling, missing data and cluster-number selection; and evaluate reproducibility, longitudinal stability and clinical translation. Available evidence indicates that unsupervised subtyping can organize multidimensional heterogeneity and generate testable biological and clinical hypotheses. Non-input biomarkers and genetic findings support aspects of construct validity, and selected cohorts report longitudinal fertility and pregnancy associations. Structural replication through local re-clustering does not establish transport of a frozen assignment model, and outcome follow-up does not establish stable cluster membership. Evidence remains limited on whether subtype labels improve outcome prediction beyond their component variables or identify differential benefit from specific treatments. We propose eight complementary validation domains, rather than a sequential hierarchy, to summarize the evidence and identify priorities for external validation, repeated assignment, incremental prediction and treatment-effect evaluation.