A metaheuristically optimised dual-stream multimodal framework for PCOS screening with clinical and ultrasound data

Abstract Polycystic Ovary Syndrome (PCOS) is a heterogeneous endocrine disorder for which reliable screening in real-world clinical settings remains challenging due to overlapping clinical manifestations, variability in hormonal biomarkers, and inconsistencies in ultrasound imaging. Conventional diagnostic approaches that rely on either clinical–hormonal indicators or ovarian ultrasound examinations in isolation often show limited generalizability when applied to heterogeneous patient populations. This study presents a dual-stream multimodal learning framework for PCOS screening that integrates real-world clinical biomarkers with ovarian ultrasound images. The framework comprises two complementary components: a clinical stream that models endocrine and metabolic patterns using hybrid feature fusion and ensemble learning, and an imaging stream that extracts multiscale ovarian morphological features from ultrasound images using a wavelet-attention–based deep learning architecture. A metaheuristic optimisation strategy is employed to jointly tune model hyperparameters and fusion-related parameters across both streams, with the aim of improving training stability and cross-dataset performance. Feature-level and score-level fusion are subsequently applied, followed by final classification using a meta-learning model. The proposed framework is evaluated on four heterogeneous datasets (PCOSUSG, GDIFR, KFHU, and MMOTU) and is compared against clinical-only, image-only, and baseline multimodal approaches. Experimental results indicate improved discriminative performance across standard evaluation metrics. In addition, SHAP-based feature attribution and Grad-CAM visualisation are used to support interpretability and clinical relevance.