Abstracts

Rationale: Epileptic spasms (ES) most often occur in the setting of West syndrome, though occasional cases occur in older children, usually without hypsarrhythmia. There is a critical need to develop and validate an objective EEG biomarker for children with ES to facilitate its rapid diagnosis and treatment. Based on the hypothesis that interictal EEG in children with ES has distinct morphological features compared to controls, we set out to train and validate a deep learning (DL) algorithm using EEG data from children with ES.

Methods: We identified 50 patients with ES confirmed by overnight video-EEG and 50 controls who underwent video-EEG to evaluate for suspected ES but were found to have a normal EEG and deemed neurologically normal. EEG data with a minimum of five-minute segments from both sleep and awake were obtained from each subject. After applying a bandpass filter of 1-50 Hz, we selected sequential 15-second windows and applied the fast Fourier transform (FFT) to each channel to construct a channel-frequency two-dimensional representation as inputs for the neural network. The disease status (case vs. control) was served as DL training labels. We trained a state-of-the-art deep neural network model, namely ResNet, using binary cross-entropy loss and validated it using five-fold cross-validation. Finally, after the network was well trained to predict the disease status, we adopt interpretability analysis (occlusion sensitivity) to extract salient interictal EEG features discriminating cases from controls.

Results: The trained model demonstrated an accuracy of 94% (five false negatives, one false positive) and a patient-wise F1 score of 0.96 in sleep EEG data. In awake EEG data, the accuracy decreased to 89% (eleven false negatives), and the patient-wise F1 score decreased to 0.87. The occlusion sensitivity analysis demonstrated fast-frequency activities (alpha-beta range) across the channels on FFT, corresponding with epileptiform discharges on EEG tracing, as the salient features of interictal EEG in children with ES (Figure 1).

Conclusions: This study suggests that a DL model, once trained, can reliably distinguish between children with and without ES, especially in sleep EEG recordings. By conducting interpretability analysis with the occlusion sensitivity, we identified salient features of interictal EEG in children with ES based on the FFT map. Further study is needed to validate these findings and enhance the accuracy of the model discriminating children with ES from control or from other types of epilepsy.

Funding: This study was accomplished with support from the Pediatric Victory Foundation, the Sudha Neelakantan & Venky Harinarayan Charitable Fund, the Elsie and Isaac Fogelman Endowment, the Mohammed F. Alibrahim Endowment, the Hughes Family Foundation, the John C. Hench Foundation, the UCLA Children’s Discovery and Innovation Institute, and UCB Biopharma.