PCOS-FedSev: Federated severity-aware PCOS diagnosis with site-invariant disentanglement and conformal guarantees

Abstract

Abstract Polycystic ovary syndrome (PCOS) is a condition that affects between 8 and 13% of women during their reproductive years. However, PCOS continues to be vastly underdiagnosed despite its prevalence. Currently available AI-based PCOS diagnosis methods are limited by their centralized nature, making them difficult to deploy across multiple sites. They are not able to provide information regarding the severity of the PCOS diagnosis. The PCOS-FedSev method is comprised of five interwoven modules such as (1) Gradient Reversal Layer-Based Site-Invariant Feature Disentanglement (SiFD) for removing the domain shift induced by the medical imaging device across multiple hospitals (2) Conditional Ordinal Regression for Neural Networks (CORN), a technique enabling joint four-level monotonic severity classification alongside binary PCOS identification (3) Federated Prototype Replay (FPR), which avoids catastrophic forgetting when incorporating a new hospital without exchanging data between clients (4) Adaptive Quality-Aware Site Weighting (AQSW) enabling robust aggregation of the contribution of clients with differing image quality and finally (5) Federated Conformal Prediction (FCP), allowing for approximately 95% marginal coverage guarantees. In experiments performed on six simulated heterogeneous hospital clients with Dirichlet α = 0.4—corresponding to realistically varying PCOS prevalence between 0 and 85% across hospitals—PCOS-FedSev demonstrated an AUC-ROC of 1.0000, an F1 score of 0.9975, a severity MAE of 0.7365, an ECE of 0.0025, and reduced communication cost by 85%. This work constitutes the first federated learning framework for PCOS identification.