Abstracts

Rationale: Mechanisms of epilepsy are poorly understood and despite any available treatment approximately 30% of patients are refractory. Over the last two decades, the evidence has been growing that in addition to epileptic spikes and discharges, high frequency oscillations (HFOs, ripples, 100-250 Hz and fast ripples, >250 Hz) are important biomarkers of the epileptogenic tissue (1-3). Those events represent a challenge to detect them due to their small amplitude. New methods of artificial intelligence such as Deep Learning (DL) neural networks can provide additional tools for automated analysis of EEG. Here we present a Long Short-Term Memory (LSTM) neural network trained to detect spikes, ripples alone and ripples riding on spikes (ripple-on-spike, RonS). Methods: We used intracranial EEG (iEEG; fs=500 Hz) from 2 independent datasets. First dataset (6 patients from the open access database ieeg.org, Mayo clinic, the MC dataset) was used for network training and testing. The second dataset (5 patients from Augusta University Health, the AUH dataset) was used for cross-institutional validation. Spectral features (relative power within traditional frequency bands) were analyzed at each time bin (0.25 s) and the initial screening for event candidates was done with the thresholding algorithm. Up to 1000 events of each class (spike, RonS, ripple and baseline) were manually selected from the candidates. The DL neural network created with MATLAB consisted of 5 layers including the bidirectional LSTM layer with hidden units varied from 50 to 200. Training was performed using random selections of 50-500 events (per class) while other 500 events (per class) were used for testing. This ‘global’ network trained on all patients from the MC dataset was then tested separately for each patient from both datasets. Results: Representative spike, RonS and ripple detected by the network are shown in Figure. The network was able to detect events of each class (not used in training) in each patient. The average values for sensitivity/specificity were: 91/99% (spikes), 92/97% (RonS) and 97/98% (ripples). We also correlated the number of events across all channels in each patient with seizure onset zone (SOZ) determined by neurologists. Spikes (corr=0.42, p<0.05) and RonS (corr=0.27, p<0.05) showed significant correlation with SOZ in all patients from the MC dataset and in 3 out of 5 patients from the AUH dataset while ripples correlated with SOZ in 4 out of 7 patients from the MC dataset and in one patient from the AUH dataset. The lower correlation of ripples with SOZ can be explained by the fact that the clinical determination of SOZ is not based on ripples which are impossible to detect without spectral analysis. Conclusions: The DL networks can be used for automated detection of epileptiform events within iEEG. Most importantly, the trained LSTM network shows generalizability detecting events with sensitivity and specificity higher than 90% in all patients tested including cross-institutional validation. The DL networks can significantly accelerate the analysis of iEEG data and increase their diagnostic value which may improve surgical outcome in patients with localization-related refractory epilepsy.References1. Jacobs J, Zijlmans M, Zelmann R et al. (2010) High-frequency electroencephalographic oscillations correlate with outcome of epilepsy surgery. Ann Neurol 67:209-202. Jobst BC, Engel J (2015) Is it time to replace epileptic spikes with fast ripples? Neurology 85:114-53. Staba RJ, Stead M, Worrel GA (2014) Electrophysiological biomarkers of epilepsy. Neurotherapeutics 11:334-46 Funding: No funding