Inter-intra feature for the complementary convolutional neural network in the effective classification of epileptic seizures

The electrical activity of the brain can be monitored using the electroencephalogram (EEG), which can be used in the detection of seizures. This paper proposes an epileptic seizure detection algorithm that uses inter-intra Head-body-tail (HBT) features. Initially, the EEG signals are subdivided into non-overlapping frames. From each frame, three regions namely head (H), body (B), and tail (T) are constructed. Intra features named intra-correlative features are extracted within a frame while the inter-HBT features namely magnitude change and zero crossing feature are extracted between adjacent HBT frames. This paper also proposes a complementary convolutional neural network (CP-CNN) which uses two parallel sections of a traditional \(1D\) $1D$ -CNN network. The complementary CNN is proposed to preserve the components that are eliminated by the traditional \(1D\) $1D$ -CNN network in the feature extraction process. The evaluation of the proposed seizure detection algorithm was done with the evaluation metrics namely accuracy, sensitivity, F1-score, and specificity with the datasets namely CHB-MIT EEG dataset where the recordings were done from 22 subjects, and the Siena EEG dataset where the recordings were performed from 14 subjects. The proposed seizure detection provides an accuracy, sensitivity, F1-score, and Specificity of \(98.17\%\) $98.17\%$ , \(94.21\%\) $94.21\%$ , \(95.79\%,\) $95.79\%,$ and \(98.25\%\) $98.25\%$ respectively for the CHB-MIT dataset. For the Siena-EEG dataset the accuracy, sensitivity, F1-score, and specificity were estimated as \(98.76\%\) $98.76\%$ , \(94.15\) $94.15$ , 95.14%, and \(98.77\%\) $98.77\%$ respectively.