A Multimodal Biosignal Fusion Framework for Epileptic Seizure Detection Using Machine Learning, Ensemble Learning, and Optimization

Abstract

Epileptic seizure detection requires reliable analysis of complex physiological changes that may not be adequately represented by a single biosignal. This study developed an optimized multimodal biosignal fusion framework integrating electroencephalography (EEG), electrocardiography (ECG), electromyography (EMG), and accelerometer (ACC) signals for automated seizure detection. Data from 120 patients were processed using data quality assurance, preprocessing, feature extraction, and patient-wise partitioning to minimize data leakage. A parameterized quadratic fusion framework was applied to capture complementary and nonlinear cross-modal interactions and evaluated using twelve traditional machine-learning and deep-learning models. Particle Swarm Optimization (PSO) was employed for joint optimization of model and fusion parameters, while SHAP and PCA supported model interpretation. In addition, ensemble-learning strategies were tested for robustness and generalizability. Full quasi-quadratic fusion achieved 98.10% accuracy, 97.40% F1-score, and 98.90% AUC, while full EEG–ECG–EMG–ACC fusion achieved 99.67% accuracy and 99.79% AUC. The smallest generalization gap (ΔAUC = 0.60 percentage points) was also obtained from the fusion-aware ensemble. In summary, the results presented here demonstrate that fundamental techniques of quadratic multimodal fusion and optimization methods in machine learning for explainable ensemble approaches, when applied together, can afford an accurate, interpretable, and well-generalized fail-safe framework for multimodal automated epileptic seizure detection.