Abstracts

Rationale: Feature extraction is the basis of machine learning techniques for analyzing electrocorticographic (ECoG) data. Hand-engineered features have been used traditionally but are limited by human creativity.  Deep learning can overcome this limitation by automating feature extraction. Methods: ECoG records from 256 patients in the NeuroPace RNS® System Pivotal Study were converted to spectrograms (1-125 Hz, 0.5 sec window) for each ECoG channel. Spectrograms were resized to 227x227x3 images that were analyzed by a pre-trained AlexNet convolutional neural network (CNN) to address two problems: (1) classification of scheduled vs long-episode ECoGs, and (2) prediction of patients who benefitted vs patients who did not benefit from treatment. For the classification problem, randomly selected images of 1000 scheduled ECoGs (baseline ECoGs stored at fixed times every day) and 1000 long episode ECoGs (ECoGs triggered by detection of long epileptiform events such as electrographic seizures) were used as targets for the CNN classifier. For the prediction problem, randomly selected images of 8000 scheduled ECoGs from those who benefitted (patients with a 50% or greater reduction in clinical seizures compared to baseline for at least 50% of time since RNS System implant, n = 142) and 8000 scheduled ECoGs from patients who did not experience a sustained improvement from pre-implant baseline (n = 18) were used as targets for the CNN classifier. Stochastic gradient descent optimizers with initial learning rates of 0.001, mini-batch size of 64 and 20 training epochs were set as training parameters for both problems. Performance of deep learning and traditional machine learning using hand-engineered features (interictal spikes, total spectral power and power in different frequency bands) were compared using a weighted k-nearest neighbor based classifier trained on the same 16,000 ECoGs from those who benefitted and those who did not benefit from treatment. An 80/20 training/testing split was used for all analyses. Results: Using test data, the CNN performed with 96.5% accuracy on the scheduled vs long episode ECoG classification problem. The prediction accuracy was 86.4% for those who benefitted vs patients who did not benefit from treatment. Hand-engineered features performed with 82.1% accuracy on the prediction problem. Visualization of 1st layer activations of the CNN shows activations along different horizontal and vertical edges suggesting that the features learned by the deep learning algorithm might be similar to the hand-engineered features used in this study. Conclusions: We have demonstrated that deep learning with a pre-trained network such as AlexNet has the potential to outperform traditional machine learning techniques for analyzing ECoG data. A deep analysis into the features and feature weights of the different layers in the trained CNN classifier may reveal features with a potential to be seizure biomarkers. Funding: None.