Diagnosis of Systemic Lupus Erythematosus (SLE) and Lupus Nephritis (LN) based on a Raman-Metabolomics multimodal re-annotated segment hierarchical fusion model.

Significant advances have been achieved in the field of Computer-Aided Diagnosis (CAD) for Systemic Lupus Erythematosus (SLE) and Lupus Nephritis (LN), particularly in the field of multimodal fusion diagnosis. The integration of Raman spectroscopy and metabolomics presents a promising direction for disease diagnosis; however, current Metabolomics-Raman fusion models still encounter challenges such as inadequate interpretability and inefficient utilization of inter-modal interaction information. To address these issues, this study proposes a dedicated multimodal fusion framework for Metabolomics-Raman data, encompassing a Metabolite-Raman re-annotation segmentation method and a Multimodal Segment Hierarchical Fusion Model (MSHFM) constructed based on this method. Through the screening of differential metabolites and the analysis of their superclasses, a mapping relationship was constructed that directly links molecular vibration modes (as identified by Raman spectral wavenumbers) to metabolic superclasses. This approach facilitates metabolically meaningful spectral segmentation, overcoming the limitations of traditional physical partitioning. The MSHFM excavates the correlations between spectral segments and differential metabolites through the Cross-Attention Pre-fusion Module (CAPM), and realizes hierarchical fusion and feature enhancement via the Sample-level Adaptive Multimodal Segment Fusion Module (Slam-SFM). The classification accuracies on the SLE-LN and HC-SLE datasets reach 93.33% and 100%, respectively. We identified key biomarkers such as uric acid and prostaglandin E1 by analyzing the Metabolomics-Raman association patterns using the attention mechanism, which provides an interpretable new method and potential biomarkers for the diagnosis and mechanistic research of SLE and LN.