A Deep Multimodal Fusion Framework Integrating SERS, Clinical Data, and Metabolomics for Noninvasive MASH Diagnosis

Weilin Wu, Jing Wang

Abstract: Metabolic dysfunction-associated steatotic liver disease (MASLD), particularly its progressive form, metabolic dysfunction-associated steatohepatitis (MASH), presents significant diagnostic challenges due to its complex etiology and heterogeneous clinical presentation. While surface-enhanced Raman spectroscopy (SERS) offers promising molecular-level sensitivity, its diagnostic accuracy is limited when used alone, hindering its suitability for clinical diagnostic applications. To address these challenges, we introduce MedFusionNet (MFN), a novel multimodal deep learning framework that integrates SERS spectra, clinical-biochemical (CB) parameters, and bile acid (BA) metabolomics for noninvasive MASH diagnosis. MFN leverages a Product-of-Experts (PoE)-based variational fusion architecture to capture intermodality interactions and latent correlations, and jointly learns a shared latent representation through a multiloss optimization strategy. When applied to a biopsy-confirmed MASLD cohort (n = 240), MFN outperformed both conventional and multimodal baseline models in predictive accuracy and robustness. SHapley Additive exPlanations (SHAP), combined with stratified 5-fold cross-validation, provides fine-grained interpretability across modalities. These findings underscore MFN’s potential as a scalable, clinically actionable framework for early MASH detection and broader applications in complex disease diagnostics.