Abstracts

Predicting TLE Laterality Using Machine Learning of Hippocampal Dynamic Functional Connectivity

Abstract number : 1.212
Submission category : 4. Clinical Epilepsy / 4B. Clinical Diagnosis
Year : 2018
Submission ID : 501101
Source : www.aesnet.org
Presentation date : 12/1/2018 6:00:00 PM
Published date : Nov 5, 2018, 18:00 PM

Authors :
Sharon Chiang, University of California - San Francisco; Emilian Vankov, Rice University, Baker Institute of Public Policy; Hsiang J. Yeh, University of California - Los Angeles; and John M. Stern, University of California - Los Angeles

Rationale: Although highly effective, surgery is not pursued in about 30% of patients suffering from temporal lobe epilepsy (TLE), primarily due to inconclusive lateralization. Functional connectivity MRI may provide a useful adjunct to the surgical workup. Recent evidence suggests that dynamic functional connectivity (dFC) may provide additional information about pathologic connections not contained in traditional static connectivity estimates. We investigate whether machine learning of hippocampal dynamic functional connectivity improves prediction of TLE laterality over traditional functional connectivity methods. Methods: Interictal resting-state fMRI was performed in 24 healthy controls and 30 patients with unilateral temporal lobe epilepsy with candidacy for anterior temporal lobe resection based on epilepsy surgery evaluation, including 19 patients with left TLE and 11 patients with right TLE. A GARCH dynamic conditional correlations (GARCH-DCC) model was used to estimate dynamic functional connectivity within the ipsilateral and contralateral hippocampus for each patient. Fast Fourier transform was applied to dynamic functional connectivity estimates. Temporal and power spectral features of dynamic functional connectivity were computed within the ipsilateral and contralateral hippocampus and used within random forests to predict TLE laterality. Five-fold cross validation was used to compare predictive accuracy in TLE lateralization based on dFC features as compared to traditional static functional connectivity estimation. Results: Five-fold cross validation using temporal and spectral features of dynamic functional connectivity within the hippocampus achieved 74% accuracy in predicting TLE laterality, with a 13% false positive rate. Utilization of dynamic functional connectivity provided significantly greater predictive accuracy than traditional estimates of connectivity relying on the Pearson correlation, which corresponded to 63% accuracy, with a 25% false positive rate. Conclusions: Dynamic functional connectivity analysis of interictal resting-state fMRI may improve the ability to lateralize TLE compared to traditional approaches of analyzing resting-state fMRI connectivity through static estimation methods. Establishment of the clinical utility of machine learning for TLE lateralization is needed and may ultimately provide a useful adjunct in the surgical evaluation for patients with TLE. Funding: None