Automatic detection of High Frequency Oscillations (80-500Hz) based on Convolutional Neural Network in Human Intracerebral Electroencephalogram.

Recently, high-frequency oscillations (HFOs) of range 80-500 Hz in electroencephalogram (EEG) recordings of epilepsy patients are considered as a reliable marker of epileptic seizure. In the present work, an automatic detection of HFOs represents an isolated peak (an `island') in a time-frequency plot based on convolutional neural network (CNN) was proposed. Initially, three patients with medically intractable epilepsy were recruited. They underwent a presurgical monitoring individually with around 54-90 channels of intracranial electroencephalograph (iEEG). Then, a specific CNN with five layers was developed with a total of 18,400 time-frequency island pictures marked with a label of either a real HFO or a false HFO. They are in the range of 80-500 Hz in the recorded iEEGs of 312 hours. Besides, over 7940 pictures including 3970 real HFO events and 3970 false HFO events except the training set were used to evaluate the performance of the current proposed method. As a result, the obtained precision of HFO events, the value of the recall, and the F1 score of the proposed CNN were found to be 94.19%, 89.37%, and 91.71%, respectively. Additionally, the automatic detection time of each HFO event is limited within 1-3 seconds. In summary, the proposed HFOs detector with deep learning would be more efficient and useful in the diagnosis of epilepsy as compared with the current manual determination of each HFOs from a long-term multichannel iEEGs recordings.