Abstracts

Rationale: The field of intracranial EEG (iEEG) processing in epilepsy diagnostics and research has undergone rapid development in the last decades. Even though the automatic tools show promise for more efficient localization of epileptogenic tissue than manually determined seizure onset zone (SOZ), their transition into the clinic is slow. Introducing signal processing algorithms and machine learning models into clinical practice can significantly improve clinical decision making and lead to increase well-being of patients. Here we have developed and tested a unified software library - the Electrophysiology Computational Module for automatic iEEG feature extraction and machine learning tasks. Methods: The unified software library for automatic localization of epileptic tissue was used to study 18 patients with the excellent postsurgical outcome (ILAE 1 & Engel IA) with >1-year follow up. We processed 30 minutes of pre-operative interictal resting state, awake iEEG recordings sampled at 5 kHz by all modules available in the library - oscillatory events and high frequency oscillation (HFO) detection algorithms, univariate feature extraction - phase-amplitude coupling, frequency-amplitude coupling, signal entropy, and signal power and bivariate feature extraction - linear correlation, phase-lag index, relative entropy and phase synchronization. Phase-amplitude coupling and frequency amplitude coupling were computed with 1-30 Hz filter signal for low frequency band and 65-180 Hz filter signal for high frequency band while the rest of the features were calculated for 8 bands (delta 1-4Hz, theta 4-8Hz, alpha 8-12Hz, beta 12-20Hz, low-gamma 20-45Hz, high-gamma 55-80Hz, ripples 80-250Hz, fast-ripples 250-600Hz). We tested the localization potential of each feature by Wilcoxon rank-sum test comparing features in resected SOZ channels and the remaining channels. The extracted features were subsequently used to train a support vector machine algorithm. The utility of the machine learning model for successful classification of resected channels was tested by calculating the receiver operating curve (ROC) and the corresponding area under the curve (AUC). Results: Oscillatory event detection showed the best performance in alpha, low gamma and high gamma bands (p<0.01). Neither of the coupling features nor signal entropy reached significance but power of the signal showed significant differences in all frequencies above alpha band (p<0.01). Linear correlation showed the best performance in theta band (p<0.01) and alpha, beta and low-gamma band (p<0.05). Phase synchrony reached p<0.05 in wide range: delta – low-gamma. Phase-lag index showed the best results in low- and high-gamma (p<0.01) and in alpha and beta band (p<0.05). Relative entropy achieved p<0.001 in all frequency bands. The trained machine learning model reached AUC of 0.838 for correct classification of resected contacts. Conclusions: The presented software library provides algorithms that might localize epileptogenic tissue from 30 minutes of interictal recording. The library can be used in research and can help transition research results into clinical practice and, thus, significantly shorten patients’ stay in hospital. Funding: Supported by the project no. LQ1605 (MEYS CR, NPU II) and Inter-Action (LTAUSA18).