Automated spectrographic seizure detection using convolutional neural networks

Non-convulsive seizures are common in critically ill patients, and delays in diagnosis contribute to increased morbidity and mortality. Many intensive care units employ continuous EEG (cEEG) for seizure monitoring. Although cEEG is continuously recorded, it is often reviewed intermittently, which may delay seizure diagnosis and treatment. This may be mitigated with automated seizure detection. In this study, we develop and evaluate convolutional neural networks (CNN) to automate seizure detection on EEG spectrograms.

Adult EEGs (12 patients, 12 EEGs, 33 seizures) from New-York Presbyterian Hospital (NYP) and pediatric EEGs (22 patients, 130 EEGs, 177 seizures) from Children's Hospital Boston (CHB) were converted into spectrograms. To simulate a telemetry display, seizure and non-seizure events on spectrograms were sequentially sampled as images across a detection window (26,380 total images). Four CNN models of increasing complexity (number of layers) were trained, cross-validated, and tested on CHB and NYP spectrographic images. All CNNs were based on the VGG-net architecture, with adjustments to alleviate overfitting.

For spectrographically visible seizures, two CNN models (containing 4 and 7 convolution layers) achieved >90% seizure detection sensitivity and specificity on the CHB test set and >90% sensitivity and 75-80% specificity on the NYP test set. The one CNN model (10 convolution layers) did not converge during training; while another CNN (2 convolution layers) performed poorly (60% sensitivity and 32% specificity) on the NYP test set.

Seizure detection on EEG spectrograms with CNN models is feasible with sensitivity and specificity potentially suitable for clinical use.