Authors :
Presenting Author: Jared Pilet, – Marquette University/Medical College of Wisconsin
Thomas Luo, BS – Medical College of Wisconsin; Chad Carlson, MD – Professor, Department of Neurology, Medical College of Wisconsin; Christopher Anderson, MD – Associate Professor, Department of Neurology, Medical College of Wisconsin; Wade Mueller, MD – Professor, Department of Neurosurgery, Medical College of Wisconsin; Sean Lew, MD – Professor, Department of Neurosurgery, Medical College of Wisconsin; Scott Beardsley, PhD – Associate Professor, Joint Department of Biomedical Engineering, Marquette University/Medical College of Wisconsin; Manoj Raghavan, MD PhD – Professor, Department of Neurology, Medical College of Wisconsin
Rationale:
Seizure onset zones (SOZs) identified from ictal intracranial recordings remain the clinical gold standard to identify targets for epilepsy surgery. However, several interictal features of ECoG signals such as epileptic spikes, high-frequency oscillations (HFOs), or focal slow activity provide valuable information about the location and extent of epileptogenic zones (EZ). Functional connectivity analysis and graph-theoretic metrics derived from them offer potential new biomarkers. The sensitivity and specificity of any method of identifying EZs based on interictal data can be expected to improve by combining information derived from multiple features. We hypothesized that machine learning using SOZs or resection zones (RZs) as the targets for training offers a viable approach for integrating information from an array of putative interictal biomarkers and assessing their value in predicting the EZ.
Methods:
Awake 30-minute samples of interictal ECoG data were retrieved for 26 patients who maintained an Engel Class I seizure outcome for at least two years after epilepsy surgery. We generated binary maps of the SOZ from clinical determinations made from ictal recordings prior to surgery. Binary maps of RZ were generated based on a visual review of the locations of intracranial electrodes relative to post-resection brain imaging. The following features of the ECoG signals were extracted from all channels using custom MATLAB scripts: epileptic spike rates and mean amplitudes, HFO rates and mean amplitudes, and power spectral density (PSD) in six frequency bands from delta to high-gamma. Functional connectivity measures were calculated for five different frequency bands in two different ways for each patient: (1) narrow-band amplitude envelope correlations (AEC), and (2) phase locking values (PLV). Two graph metrics, node strength (NS), and eigenvector centrality (EC) were calculated from these connectivity matrices for all frequency bands. This yielded a total of 30 features for each of the 2176 electrode locations. Four-fold cross-validation was employed to evaluate the prediction of SOZs and RZs using a support vector machine (SVM).
Results:
For predicting the SOZ, an SVM classifier yielded an AUROC of 0.90 using all 30 features. When using graph-theory-based features alone, SOZ was classified with an AUROC of 0.85. This value surpasses the results observed using spike and HFO features alone (0.70) or PSD alone (0.78). The same classification method predicted the RZ with an AUROC of 0.90 using all features. When using graph-theory-based features alone, SOZ was classified with an AUROC of 0.85. This value also surpasses the results observed when using spike and HFO features alone (0.64) or PSD alone (0.76).
Conclusions:
Our results show that machine learning provides a robust approach for integrating information from multiple putative biomarkers to identify epileptogenic cortices. They also demonstrate that features derived from interictal ECoG functional connectivity carry valuable additional information relevant to identifying epileptogenic zones.
Funding: None.