Automated Detection of High Frequency Oscillations in Intracranial EEG Using the Combination of Short-Time Energy and Convolutional Neural Networks

High-frequency oscillations (HFOs) of 80~500 Hz in the intracranial electroencephalogram (iEEG) recordings are considered as a reliable marker for epileptic location. However, a significant challenge to the clinical use of HFOs is due to the time-consuming procedure of visually identifying them. A new methodology is presented in this paper for the automated detection of HFOs based on their 2D time-frequency map employing the short-time energy (STE) estimation and the convolutional neural network (CNN) classification algorithm. The effectiveness and usefulness of the proposed method are evaluated using the clinical iEEG data acquired from five patients (28.4 ± 13.0 years) with medically intractable epilepsy. The proposed methodology presents the following significant advantages: 1) compared with the recently reported HFOs detector based on the CNN using only the 1D temporal EEG signal, the proposed method achieves a higher accuracy using the deep CNN classifier on 2D time-frequency map of HFOs, of which the evaluated sensitivity and false discovery rate (FDR) for identifying ripples are 88.16% and 12.58%, respectively, and the corresponding sensitivity and FDR are 93.37% and 8.1% for detecting fast ripples, respectively; 2) it is capable of automatically extracting the shared features of HFOs events of different patients and would be much robust, unlike other automated methodologies proposed in the literature where the characteristics of HFOs were extracted manually on the basis of researchers' knowledge, which, probably, is prone to observer bias; and 3) with the proposed STE estimation, all suspicious ripples and fast ripples could be initially found out and transformed into time-frequency map for subsequently CNN-based classification, rather than transforming and classifying the raw data, thus requiring a lower computational resource. In addition, the time occurrence of each transient event of the HFOs can be identified to be potentially useful for further seizure analysis. In conclusion, this automated detection of the HFOs combing the STE and the CNN could allow analyzing large amounts of data in a short time while assuring a relatively higher accuracy and, thus, would potentially serve to provide a clinically useful tool.