SoleFusion-Net: an explainable multimodal deep learning framework for diabetic foot syndrome classification in type II diabetes mellitus.

Early detection of diabetic neuropathy remains a challenge due to reliance on subjective clinical assessments. We propose SoleFusion-Net, an explainable multimodal deep learning framework that integrates plantar pressure images with structured clinical data through an late-fusion dual-branch architecture. The image branch employs convolutional layers to extract biomechanical pressure features, while the tabular branch processes clinical and demographic variables. Learned representations are fused for joint classification. Extensive explainability techniques, including Grad-CAM, SHAP, surrogate decision trees, and prototyped critique analysis, were implemented to enhance transparency and clinician trust. Evaluated on 504 patients stratified into mild, moderate, and severe neuropathy, SoleFusion-Net achieved 83% validation accuracy with AUCs of 0.962, 0.892, and 0.933 across the three classes. Explainability analyses identified vibration perception threshold and monofilament scores as critical predictors, while Grad-Cam highlighted spatially significant plantar regions. SoleFusion-Net showcases the feasibility of adjusting the standard multimodal fusion process and the use of various XAI methods to achieve the goal of diabetic foot syndrome classification. Therefore, this solution ensures a trustworthy help about risk stratification. The framework provides a modular and explainable AI model for integrating multimodal biomedical data, with a structure that is potentially generalizable to other settings pending external validation.