Epileptic Seizure Detection by Transfer Learning Considering Ictal-like Non-ictal Signals in Electroencephalogram

Epileptic seizure detection based on the electroencephalogram (EEG) is essential in the diagnosis and treatment of epilepsy patients. However, there is a fact that affects the detection but has so far been ignored. That is, some non-ictal EEG signals are similar to the signals of ictal periods. Although the public dataset is artifact free, we also identified this kind of signals. Such ictal-like non-ictal signals tend to be misclassified as seizure signals, thus have a negative effect on seizure detection. Considering existence of these signals, we propose a method that recognizes and utilizes them to benefit the classification. First, the ictal-like non-ictal signals are recognized from non-ictal signals by a borderline-recognition algorithm calculating number of majority class samples in k nearest neighbors. Then this kind of samples is enhanced in the training of the classifier. A feature of time-frequency domain, i.e., the 2- d time-frequency diagram of the EEG, is computed from the original 1-d EEG signal and selected to be features used in the classification. A deep learning model that is previously well trained on ImageNet in image recognition tasks is transferred here for solving the time-frequency diagram classification task. Our experiments on the Bonn database show that the overall accuracy of seizure detection is increased by 1.79%. Moreover, without consideration of the ictal-like non-ictal signals, 25% of the EEG signals are misclassified as seizures. In comparison, the proposed method reduces this percentage significantly to 7.1%.