Abstract Details

Title Functional Connectivity for Disease Classification in Epilepsy: A Machine Learning Approach

Topic Epilepsy/Clinical Neurophysiology (EEG)

Presentation(s) S10 - Epilepsy and Clinical Neurophysiology (EEG) 2 (4:08 PM-4:16 PM)

Poster/Presentation
Number 002

Background Functional connectivity analysis has moved to the forefront of human brain research. Epileptic networks can be studied using functional connectivity. Machine learning methods are increasingly used in analysis of complex data.

Objective

To demonstrate if functional connectivity (FC) analysis measures can distinguish patients with epilepsy from controls and to lateralize the seizure onset zone in patients by using a supervised machine learning method.

Design/Methods

Resting state fMRI from 85 patients evaluated for epilepsy surgery (59 adults) and 33 controls (17 adults) were retrospectively selected from three centers. The connectivity adjacency matrix, mean, and degree of FC for each parcel were derived from parcel-wise correlation analysis. After feature selection using Lasso method, support vector machine (SVM) classifiers were trained to distinguish epilepsy from control (primary analysis), and hemispheric onset in epilepsy patients (secondary analysis). Classification accuracy and feature weights were evaluated.

Results

Epilepsy classification reached accuracy of 100% in an unseen test group using the connectivity matrix. Connections to default mode network, and between its parcels showed the highest weights for classification. The SVM using mean and degree of FC models reached accuracies of 93.5 and 80.8%, with parcels in visual and somatomotor networks showing the highest weights. Focus lateralization reached average accuracy of 95.5% for connectivity matrix, 80.3 for mean, and 89.5% for degree of FC.

Conclusions Alterations in FC in epilepsy can be detected and used by machine learning approaches to distinguish epilepsy patients from controls, and further lateralize focal epilepsy. Future studies including larger heterogeneous groups of patients will serve to develop seizure focus localization tools and predictors of treatment derived from fMRI data.

Authors/Disclosures
Taha Gholipour, MD (UC San Diego Health - Comprehensive Epilepsy Center) PRESENTER	Dr. Gholipour has nothing to disclose.
	No disclosure on file
	No disclosure on file
William D. Gaillard, MD (Children'S National Hospital)	The institution of Dr. Gaillard has received research support from all federal or foundation grants, NINDS, NIDCD, NICHD, NSF ,PERF.