A Multimodal Swin-LSTM Fusion Model with Explainable AI for Brain Stroke Detection and Classification

Brain stroke remains a major cause of death and disability worldwide. While most automated stroke detection methods rely on a single modality, this study presents a multimodal approach combining standardized clinical data, Brain Magnetic Resonance Imaging (MRI), and retinal Optical Coherence Tomography (OCT). A lightweight Shifted Window (Swin) Transformer and Bidirectional Long Short-Term Memory (BiLSTM) architecture is proposed for tri-class stroke classification (hemorrhagic, ischemic, normal). Clinical features are encoded via a multilayer perceptron, while Swin backbones extract discriminative MRI and OCT embeddings. The LSTM performs late fusion on concatenated features, with SMOTE addressing class imbalance. Explainability is achieved using Grad-CAM for imaging and SHAP for clinical data. The fused model attains 93.52% testing accuracy, outperforming unimodal baselines (clinical: 90.0%, MRI: 90.5%, OCT: 89.0%) and strong vision backbones such as DenseNet201 and MobileNetV3. The proposed interpretable Swin–LSTM framework demonstrates superior accuracy and generalizability, offering a robust foundation for automated multimodal stroke screening and subtype classification.