Abstracts

Rationale: The emerging evidence suggests that in addition to epileptic discharges, high frequency oscillations (HFOs) are important biomarkers of the epileptogenic tissue.^1-2 The interaction of oscillations across different time scales is revealed as cross-frequency coupling (CFC) representing a high-order structure in the functional organization of brain rhythms. New methods of artificial intelligence such as deep learning (DL) neural networks can provide powerful tools for automated analysis of EEG including the analysis of the CFC patterns which are likely to reflect different functional states of the brain. Here we present a Stacked Sparse Autoencoder (SSAE) trained to recognize absence seizure activity from the preceding preictal activity based on the CFC patterns within scalp EEG.

Methods: We used scalp EEG records (fs = 256 Hz) from the Temple University Hospital database (the TUSZ corpus³) with seizures annotated by neurologists. All absence seizures (n = 94) from 12 patients were taken into analysis along with segments of preictal activity selected randomly so that to make the total duration of preictal and ictal activity comparable. Half of the records was selected randomly for network training and the second half was used for testing. Power-to-power coupling was calculated between all frequencies 2 to 120 Hz pairwise using the EEGLAB toolbox. The resulting CFC matrices for all EEG segments were used as training or testing inputs to the autoencoder.

Results: The SSAE network created with MATLAB (v. R2021b) consisted of two encoder-decoder networks and the softmax layer with two outputs for binary classification "seizure vs. preictal." During training, the network with L2 and sparsity regularizers achieved a squared error smaller than 10^-2 with 400 iterations. The trained network was able to recognize preictal and ictal segments (not used in training) with sensitivity of 97.9%, specificity of 90.0%, and the overall accuracy of 96.5%. Our post hoc analysis showed that the major difference between preictal and ictal activity was due to an increase in the power-to-power coupling within beta-gamma bands (13-120 Hz) during seizures.

Conclusions: Our results provide evidence that the SSAE neural networks can be used for automated detection of seizures within scalp EEG. Importantly, the trained SSAE network showed generalizability detecting seizures with sensitivity and specificity at or higher than 90% in all patients tested. The deep learning networks can significantly accelerate the analysis of EEG data and increase their diagnostic value.
_x000D_ References:_x000D_ 1. Jacobs J, Zijlmans M, Zelmann R, Chatillon CE, Hall J, Olivier A, Dubeau F, Gotman J. High-frequency electroencephalographic oscillations correlate with outcome of epilepsy surgery. Ann Neurol. 2010;67:209-220._x000D_ 2. Staba RJ, Stead M, Worrel GA. Electrophysiological biomarkers of epilepsy. Neurotherapeutics. 2014;11:334-346._x000D_ 3. Shah V, Von Weltin E, Lopez S, McHugh JR, Veloso L, Golmohammadi M, Obeid I, J. Picone. The Temple University Hospital seizure detection corpus. Front Neuroinform. 2018;12:83.

Funding: NIH Grant MH123192 to A.M.