Authors :
Presenting Author: Jan Cimbalnik, PhD – International Clinical Research Center, St. Anne's University Hospital, Brno, Czech Republic
Petr Klimes, PhD – Institute of Scientific Instruments, The Czech Academy of Sciences, Brno, Czech Republic; Vojtech Travnicek, MSc. – International Clinical Research Center, St. Anne's University Hospital, Brno, Czech Republic; Martin Pail, PhD – 1st Department of Neurology, Faculty of Medicine, Masaryk University, Brno, Czech Republic; Milan Brazdil, PhD – 1st Department of Neurology, Faculty of Medicine, Masaryk University, Brno, Czech Republic
Rationale:
Localization of epileptogenic tissue in pharmacoresistant epileptic patients who undergo invasive EEG remains a challenge and requires prolonged monitoring in order to record seizures. Research of interictal (seizure-independent) electrophysiological biomarkers has shown promise to accelerate the delineation of seizure generating tissue, however, none of the biomarkers reached a sufficient localization potential to penetrate into clinical practice. Integration of multiple biomarkers using machine learning techniques remains highly unexplored. In this study we investigated the potential of multiple machine learning models for localization of epileptogenic tissue.
Methods:
We selected 20 patients out of 135 consecutive patients who underwent stereo-EEG monitoring (N=3007 channels). The selection criteria were: determined seizure onset zone, determined resected area, and positive surgical outcome (Engel IA). In 30 minute interictal resting state iEEG (sampling frequency 5kHz), we detected interictal spikes, high frequency oscillations, and computed functional connectivity measures represented by relative entropy and linear correlation. The connectivity measures were computed for different frequency bands (1-4Hz, 4-8 Hz, 8-12 Hz, 12-20 Hz, 20-55 Hz, 65-80 Hz, 80-250 Hz, 250-600 Hz). The features were preprocessed by transforming their distribution to gaussian distribution. To reduce the number of features we used a mutual information metric. Subsequently, we trained and tested six machine learning models with varying parameters and number of selected features (3-8) using a leave-one-out cross validation with weighted f1 scoring. To determine the best model for the localization task we ranked the models based on weighted f1 score. We used three definitions of the epileptogenic tissue around implanted contacts - seizure onset zone, resected area and an overlap between seizure onset zone, and the resected area.
Results:
Irrespective of the number of selected features and the localization target, the ensemble machine learning models (gradient boosting; f1=0.974 and random forest; f1=0.974) outperformed the simpler models (support vector machine; f1=0.941 and logistic regression; f1=0.936). The models designed for outlier detection (single class SVM and isolation forest) failed in the classification task with f1 < 0.1.