Temporal Convolutional Networks Enhanced RNN Models for EEG Signal Classification

Abstract

This study systematically compares four deep learning architectures for classifying EEG signals. LSTM and GRU models, as well as TCN + LSTM and TCN + GRU models, are evaluated under the same conditions of data augmentation, hyperparameterization and early stopping.  The data consisted of recordings with a sampling frequency of 173.61 Hz, categorized into normal, interictal and ictal classes. Normalization and time domain data augmentation strategies were used in the preprocessing steps. The results show that hybrid structures provide significant superiority. LSTM produced 86.67 percent accuracy, 0.8646 macro F1, 0.7908 Matthews correlation coefficient, and 0.9699 AUROC. GRU achieved 81.33 percent accuracy, 0.8220 macro F1, 0.7083 Matthews correlation coefficient and 0.9664 AUROC. TCN + LSTM performed more strongly with 92.00 percent accuracy, 0.9221 macro F1, and 0.9953 AUROC. The highest success was achieved with TCN + GRU. This model achieved 96.00% accuracy, 0.9610 macro F1 score, 0.9375 Matthews correlation coefficient and 0.9993 AUROC. In probabilistic calibration, the best value was obtained for TCN + GRU with a log loss of 0.0982 and a Brier score of 0.0426. In terms of inference efficiency, GRU is the fastest model with an average of 0.39 milliseconds. The hybrid models meet the real-time usage threshold with a delay of about 7.6 milliseconds. The results show that the TCN + GRU architecture is successful in achieving a balance between accuracy, calibration and latency.