Phenomapping Polycystic Ovary Disorder Using Self Supervised Representation Learning and Deep Clustering of Clinical Data

Abstract: Poly cystic Ovary Disorder is one of the most prevalent endocrine conditions affecting women of reproductive age, yet its clinical presentation varies widely across reproductive metabolic and hormonal domains. Current diagnostic frameworks and computational models often simplify PCOD into a binary classification present or absent ignoring its intrinsic heterogeneity. This oversimplification hinders precision medicine and the development of tailored treatment plans. This paper introduces a novel framework that combines self-supervised learning (SSL) and deep clustering to uncover hidden PCOD phenotype directly from unlabeled clinical records. Contrastive SSL approaches such as Sim CLR and MoCo are used to learn patient embeddings which are subsequently clustered using Deep Cluster and SwAV. The resulting clusters are validated against clinical indicators including body mass index (BMI) hormonal ratios and insulin resistance measures. Experiments revealed three distinct clusters: mild PCOD with near normal parameters moderate PCOD with hormonal irregularities and severe PCOD with metabolic risks. The framework was deployed in a Django based clinical platform providing real time cluster assignments visual analytics and patient level reports. By moving beyond binary classification this work demonstrates the potential of SSL driven phenomapping to enable precision gynecology supporting individualized treatment strategies and advancing scalable clinical decision support systems.

Keywords: Polycystic Ovary Disorder, self-supervised learning, contrastive learning deep clustering, clinical phenotyping, decision support.