Tensor decomposition-based feature extraction for EEG signals in epileptic seizure detection

Epilepsy monitoring based on electroencephalogram (EEG) signals is important for detecting pre-seizure states, but multidimensional EEG representations can lead to high-dimensional feature spaces and increased computational cost. This work proposes two canonical polyadic decomposition (CPD)-based approaches for extracting compact features from spectral-temporal-spatial EEG tensors obtained with a continuous wavelet transform (CWT). The proposed approaches use the estimated low-rank factor matrices to preserve physically interpretable information while reducing the number of features used to train machine learning (ML) classifiers. Experimental results show robust classification performance for inter- and pre-ictal classification, with performance metrics of up to approximately 99%, while substantially reducing the feature-space dimensionality under patient-specific variability.