Abstracts

Rationale: Comprehensive analysis of high-frequency oscillation (HFO) rates and characteristics over multiple conditions and behavioral states across days to months requires automated HFO detection algorithms with high sensitivity and specificity. In terms of performance, Convolutional Neural Networks (CNNs) have recently exceeded traditional machine learning algorithms in a wide range of biomedical applications, and do not require a priori feature selection. We develop and train a deep CNN to identify HFOs in intracranial EEG (iEEG) recordings from patients undergoing presurgical iEEG monitoring. Methods: We divided the time series EEG data into three-second epochs from individual EEG channels into frequency bands of width 10 Hz within the range of 0-800 Hz with 5 Hz overlap in frequency ranges using a 3rd order Butterworth bandpass filter. We calculated the Hilbert transformation of the filtered data and computed the z-score of the absolute values. We then passed the 2-dimensional matrix through the CNN algorithm with 20-convolution filters of size 5x5 matrix elements followed by a second and third layer of convolution with 50 convolution filters. At first and second layers, we used a rectified linear units (ReLU) function at each neural node to calculate the node’s output and applied a 20% dropout, max-pooling, and L2 regularization techniques to reduce the dimensionality of the layer’s input data and to limit overfitting. However, in the third layer, a Sigmoid function was used as an activation function for binary classification. We used 80% of the data for training and the remaining 20% for validation. The performance of the CNN algorithm was evaluated using data segments from the two additional patients not used in algorithm training. Results: We identified 8578 segments of iEEG data from 6 patients undergoing epilepsy invasive presurgical EEG monitoring at a sampling frequency of 5000 Hz. For training and validation, we identified 1378 segments, half of which were gold standard HFOs annotated by expert epileptologists and other half non-HFOs from 4 patients. In total 7200 data segments were identified in the two held out patients. By selection of optimal hyperparameters (learning rate 0.001, number of epochs 32) we achieved an average sensitivity and specificity of 95% and 92% respectively. Conclusions: In this study, we presented an efficient detector fueled by CNN to detect HFOs. This detector may assist clinicians in localization of the epileptogenic zone in the future. However, further analysis will be done on an expanded set of expert-annotated data to provide rigorous statistical validation. Funding: This study was funded by NIH-R01NS092882, Czech Republic Grant agency (P103/11/0933), European Regional Development Fund – Project FNUSA – ICRC (CZ.1.05/1.1.00/02.0123), and a gift from Mr. and Mrs. David Hawk.