CNN with double-side weighted visibility graph for automated classification of high-frequency oscillations in epilepsy

Abstract

High-frequency oscillations (HFOs) in intracranial EEG are deemed clinically significant biomarkers for localization of epileptogenic focus. However, the identification of HFOs historically relies on visual observation which is laborious and invariably prone to inaccuracy. Here, we provided a new strategy that combines double-side weighted visibility graph (dWVG) and convolutional neural network (CNN) to automatically detect HFOs. The effectiveness of the proposed method was confirmed using stereotactic electroencephalography (SEEG) data recorded from 4 epilepsy patients, comprising 1880 HFO fragments and 1880 non-HFO fragments. In addition, the dWVG-based approach was also compared with other graph-based (i.e., time–frequency graph) methods. Results showed that the dWVG approach outperformed the time–frequency diagram and weighted visibility graph (WVG) with better classification accuracy a computing speed. By combining dWVG and CNN based on signal features instead of LeNet-like CNN, the best classification performance was obtained with an accuracy of 95.52%. These results indicate that the proposed approach plays a significant role in HFO classification, which may facilitate the localization of epileptogenic focus and provide more effective treatment for patients.