Multi-Omics and Clinical Data Integration for Enhanced Early Detection of Polycystic Ovarian Syndrome (PCOS) Using Explainable Deep Learning A Bioinformatics Approach

Abstract

Polycystic Ovary Syndrome (PCOS) is a prevalent and complex endocrine disorder affecting up to 10% of women of reproductive age. Early diagnosis remains challenging due to its heterogeneous clinical manifestations and the limitations of existing diagnostic approaches. Conventional methods that rely primarily on clinical markers often fail to detect subtle molecular abnormalities, while standalone genomic analyses lack essential clinical context, leading to delayed diagnosis and suboptimal treatment strategies.

To address these challenges, this study proposes a novel multi-modal diagnostic framework that integrates omics data with clinical biomarkers using a hybrid deep learning architecture. Gene expression datasets obtained from the Gene Expression Omnibus (GEO), specifically GSE5090, GSE54248, and GSE54250, are combined with key clinical indicators to enhance diagnostic precision. The proposed architecture employs parallel processing pathways: omics data are modeled using Graph Neural Networks (GNNs), including GATConv and GCNConv layers, to capture complex gene–gene interactions, while clinical biomarkers are processed through feed-forward neural networks. These pathways are fused using an adaptive weighting mechanism guided by clinically significant indicators such as Anti-Müllerian Hormone (AMH) levels.

To ensure robustness across heterogeneous datasets, domain adaptation techniques are applied to mitigate batch effects, and transfer learning with pre-trained encoders is utilized to improve model generalizability. The proposed system achieves a cross-validated mean Area Under the Curve (AUC) of 0.653 (±0.116). Optimized threshold tuning further enhances performance, yielding a precision of 0.975, recall of 0.870, and overall classification accuracy of up to 95%. Explainable AI techniques provide transparent insights into model predictions, consistently identifying AMH levels, β-HCG measurements, and derived ratios as key predictors. This integrative approach not only improves PCOS diagnostic accuracy but also offers valuable insights into its underlying molecular mechanisms.