Deep learning-based EEG seizure classification using comparative time–frequency spectral representations

The automated prediction and classification of seizures using electroencephalography (EEG) remains a challenging task due to the inherent complexity and significant inter-subject variability observed in seizure patterns. To address these challenges, we explored the use of two-dimensional time-frequency representations of EEG signals to enhance the effectiveness of seizure type classification. In this study, we employed EEG data corresponding to background, focal, and generalized seizures, sourced from the Temple University Hospital (TUH v2.0.1) dataset. The EEG signals were first pre-processed and segmented. Then, each segment was transformed into a two-dimensional time-frequency image using three techniques: Short-Time Fourier Transform (STFT), Mel-spectrogram, and Continuous Wavelet Transform (CWT). We developed and evaluated nine deep learning architectures for seizure classification: AlexNet, VGGNet, configurable Convolutional Neural Network (cCNN), configurable Recurrent Neural Network, configurable Long Short-Term Memory, Convolutional Neural Network–Long Short-Term Memory, Convolutional Neural Network–Bidirectional Long Short-Term Memory, Residual Bidirectional Long Short-Term Memory, and Swin Transformer. All models were assessed using nested 10-fold cross-validation and evaluated across standard performance metrics such as accuracy, precision, recall, F1-score, false positive rate, and area under the curve, along with standard deviation and 95% confidence interval. The obtained experimental results showed that CWT-based representations with cCNN achieved an accuracy of 77.22% ± 5.68% with a 95% CI of [73.32, 80.55], while Mel spectrograms with AlexNet yielded 79.83% ± 2.81% with a 95% CI of [78.20, 81.41]. In conclusion, the combination of Mel-based time-frequency features with the AlexNet and CWT with cCNN models proved to be most effective for the classification of seizure types. These results suggest that deep learning and time-frequency analysis can contribute to the development of reliable EEG-based seizure classification systems.