Automatic epileptic EEG classification based on differential entropy and attention model

In epilepsy electroencephalogram (EEG) analysis, clinicians usually interpret EEG page by page, which is time-consuming and brings heavy workload. This paper proposes a novel automatic epileptic EEG classification approach based on differential entropy and attention mechanism, aiming at designing a short-term epileptic EEG classification model with high accuracy and good generalization performance. Firstly, the original EEG recordings are decomposed into five sub-frequency bands which approximately obey the Gaussian distribution. Afterward, a improved attention model framework considering both row and column attention with a shallower VGGNet (AttVGGNet-RC) is put forward as the classifier. Finally, non-patient specific method is employed to evaluate the performance with pre-tuned hypermeters. With 8-fold data, the proposed model yielded 77.33 ± 2.91% sensitivity, 86.67 ± 3.70% specificity and 82.00 ± 1.43% accuracy, and accuracy was increased by 5.34%, 8.99%, 26.24% and 4.47% respectively compared with multi-layer perceptron (MLP), extreme learning machine (ELM), support vector machine (SVM) and Long Short-Term Memory (LSTM). With 10-fold shuffled data, the improved attention model yielded 93.84 ± 0.63% sensitivity, 95.84 ± 0.74% specificity and 95.12 ± 0.20% accuracy, and the accuracy was 1.34%, 16.29%, 27.12% and 8.24% higher than MLP, ELM, SVM and LSTM respectively. The experimental result showed that the attention model achieved high classification accuracy with low standard deviation as well as good generalization performance. Furthermore, compared with state-of-art epilepsy analysis system, the proposed approach also show better performance. Therefore, this study has significant clinical application value in epilepsy analysis.