XAI-PCOSNET: A FAITHFULNESS-VERIFIABLE MULTI OMICS DEEP-LEARNING FRAMEWORK FOR EXPLAINABLE POLYCYSTIC OVARY SYNDROME DIAGNOSIS

Abstract

Polycystic ovary syndrome (PCOS) affects a large fraction of women of reproductive age and is frequently diagnosed late, because early molecular signals are overlooked in favour of the classical clinical and ultrasound criteria. Machine-learning studies of PCOS typically read a single data layer and rarely test whether their explanations are faithful to the model, which limits clinical trust. We present XAI-PCOSNet, a modular deep learning framework that integrates multiple data modalities through modality-specific encoders, a graph-attention branch over a gene-interaction graph, and an interpretable cross-attention fusion layer that reports a weight for each modality. The central methodological contribution is a validation protocol that measures explanation faithfulness rather than assuming it: on a controlled synthetic multi-omics cohort with planted informative features, Integrated Gradients and SHAP recover 29 of 30 transcriptomic drivers (97%) and 6 of 8 hormonal drivers (75%) at a surrogate fidelity of 0.94. We then validate the framework on a real, public clinical PCOS cohort of 541 patients, where it attains the highest discrimination of all models tested (ROC-AUC 0.966, PR-AUC 0.947) with 89.7% accuracy, competitive with tuned gradient-boosting and kernel baselines; on this real cohort the attribution layer independently recovers 80% of the clinically established PCOS drivers, including follicle counts, hirsutism, skin darkening, weight gain, and irregular cycles, and the learned modality weights concentrate on the imaging and symptom modalities that clinicians actually use. The framework supports good health and well-being (SDG 3), women’s health and gender equality (SDG 5), and a reusable, verifiable computational methodology for precision medicine (SDG 9). We report negative results openly, including where classical baselines lead on fine-grained severity and where single-nucleotide-polymorphism attribution is weakest, and we release the complete implementation for reproduction.