Rationale: The objective of this research was to develop a cross-patient deep learning model that classifies seizure and non-seizure electrographic records (ECoGs) from the NeuroPace RNS® System. Such a model could potentially reduce the burden of manually reviewing large ECoG datasets. Since electrographic seizures vary from patient to patient, deep learning models that learn complex patterns directly from raw data may be more suitable for developing cross-patient seizure classifiers compared to rule or feature based methods. Methods: Data from 113 randomly selected patients (from 256 patients in the RNS System clinical trials) were used for training, validating and testing an electrographic seizure classifier. In the selected patients, 130,000 ECoG channels were manually labeled as electrographic seizures or non-seizures. Five folds of training, validation and test data were created. Each fold contained data from 72 patients for model training, 18 patients for model validation and 23 patients for model testing. Every patient was held-out in the test dataset at least once. ECoG channels were converted to three-color-channel RGB spectrograms using the jet color map and saved as 224x224x3 PNG images using the matplotlib.pyplot.specgram function (Python 3.5, Matplotlib) with window size 256 and step size 128. Five convolutional neural network (CNN) models with increasing complexity were trained on the datasets: three CNNs with six, seven and 12 layers, and two deeper ResNet architectures with 18 and 50 layers. A minimal incremental validation accuracy of 0.1% over 10 consecutive training epochs was used as a condition for early stopping of training. The trained models were tested on ECoG channels from held-out test patients. Results: The classification performance of the models increased with the complexity of the model trained. The deepest ResNet50 model performed best on ECoG channels from held-out patients with a class-balanced accuracy (CBA) of 95.7% and an F1 score 94.3%, while the shallowest 6 layer CNN performed worst (CBA=87.9%, F1=84.1%). Analysis of the other models revealed that the 12-layer CNN (CBA=94.9%, F1=93.4%) is likely the point of diminishing returns with respect to model complexity. Error analysis revealed some clear shortcomings. Models occasionally underperformed when the electrographic seizures were faint or short (< 10 seconds) and in the presence of noise artifact. Conclusions: Deep-learning models can be used to automate the task of identifying seizure ECoG records. Healthcare tools built using these models have the potential to significantly reduce the physician’s burden of reviewing EEG data for epilepsy patients. This research provides a foundation for the future development of neuromodulation devices that incorporate neural network models for seizure detection. Funding: Please list any funding that was received in support of this abstract.: None.
