Explainable multi-modal approach for uncovering key predictors of stroke-risk from ECG, EMG, blood pressure, and respiratory signals

Accurate and timely stroke-risk prediction is necessary to help patients at risk take guided measures, as stroke remains a leading cause of death and long-term disability worldwide. While there are several developed stroke-risk prediction models using various bio-signals, it remains unclear which signal or signal feature carries the most information towards stroke. Additionally, respiratory signals have not been included in these models, despite research showing their relation to stroke disease. We address these gaps by developing an explainable multi-modal stroke-risk prediction model that integrates respiratory signals (carbon dioxide (CO[Formula: see text]), and respiration flow) alongside blood pressure (BP), electrocardiogram (ECG) and electromyography (EMG). We developed a single perceptron model which achieved a prediction accuracy of 84.91% on a dataset of 64 subjects, outperforming state-of-the-art machine learning (ML) and deep learning (DL) methods. Explainable Artificial Intelligence (XAI) techniques, namely LIME, SHAP, and Anchors, were applied to interpret the model's decisions. The Model Explanation Metric Consensus (MEMC) XAI evaluation metric revealed SHAP as the most reliable explainer, which identified CO[Formula: see text]-derived features as critical predictors. The findings demonstrate the overlooked value of respiratory signals in stroke-risk prediction, as well as the importance of explainable multi-modal approaches in advancing stroke-risk prediction, and enabling clinicians to better understand and trust AI models for improved patient care.