Multimodal Classification of Epileptic Seizures based on Systematic Features Extraction from Brain and Motoric Signals

Epileptic seizures result from abnormal brain activity and involve both electrical and biomechanical signals. Accurate seizure classification is essential for clinical decision-making, yet conventional diagnostic methods, including self-reports and video monitoring, are limited in detecting seizure types. To overcome these limitations, this study investigates a multimodal approach combining electroencephalography (EEG), surface electromyography (sEMG), and inertial measurement unit (IMU) sensors. A synchronized compact wireless system was developed to capture the modalities, ensuring precise recording and analysis. A systematic signal processing pipeline was applied, including artifact removal, feature extraction, selection, evaluation, and machine learning-based classification. First, each modality was tested individually to assess its potential in seizure classification. The results revealed that a single modality was insufficient, with a maximum accuracy of 94%, highlighting the challenge of seizure similarity. To further explore multimodal classification, validation was conducted in a hospital setting. The results demonstrate that using independent component analysis (ICA) for preprocessing, feature selection techniques based on radar plots, distance metrics, and Big O notation, combined with the XGBoost classifier, led to a classification accuracy of 99%. These findings confirm that EEG, sEMG, and IMU complement each other, significantly enhancing seizure classification.