Abstracts

Rationale: Epilepsy is one of the most common neurologic disorders, affecting 1.5-2.3 million people in the U.S. and 50 million people worldwide.¹ Non-invasive methods of lateralization or localization of the seizure onset zone are crucial for diagnosis, disease stratification, and of critical interest in the preoperative workup for epilepsy surgery. Existing non-invasive methods often do not provide conclusive lateralization or localization. Non-invasive techniques that can localize the seizure-onset zone without the need for more invasive methods, such as intracranial EEG, remain in great need and could significantly improve patient outcomes and decrease socioeconomic burden.

Methods: 2132 healthy control participants and 32 pre-operative temporal lobe epilepsy (TLE) patients from ongoing studies at Washington University were studied (table 1). All participants and patients underwent structural and resting-state functional MRI (rs-fMRI). Only healthy control data were used to generate training samples for a deep three-dimensional convolutional neural network model. Synthetic noise and permutations of rs-fMRI were injected into randomly specified convex regions of rs-fMRI from healthy control participants. The specified regions sampled the hippocampus, amygdala, insula, and temporal gyri bilaterally, based on a previous report that tied abnormal functional connectivity in those regions to TLE laterality at the group-level.² Approximately 400,000 training images provided labels for left and right hemispheric synthetic seizure-onset zones. The model was then trained to classify the hemisphere containing synthetic noise. Finally, the model was tested on TLE patient rs-fMRI data to assess its performance for detecting biological seizure-onset zones.

Results: The model classified the hemisphere containing synthetic noise with 96% accuracy. It classified the hemisphere of TLE in individual patients with 87.5% accuracy. Figure 1 details additional training results.

Conclusions: Non-invasive techniques capable of localizing the seizure-onset zone could significantly improve post-surgical outcomes in patients with intractable epilepsy. In this work we have demonstrated the ability of a deep learning network to identify the correct hemisphere of the seizure onset zone at the individual level in TLE patients using resting-state fMRI data. Critically, the model was able to make accurate predictions after training exclusively on healthy control data augmented with synthetic noise. This data-driven approach represents a novel noninvasive technique of seizure lateralization that could improve preoperative surgical planning and patient outcome.

1. Ivanova JI, Birnbaum HG, Kidolezi Y, Qiu Y, Mallett D, Caleo S. Economic Burden of Epilepsy among the Privately Insured in the US Pharmacoeconomics. 2010;28:675-685.
2. Maccotta L, He BJ, Snyder AZ, Eisenman LN, Benzinger TL, Ances BM, et al. Impaired and facilitated functional networks in temporal lobe epilepsy Neuroimage Clin. 2013;2:862-872.

Funding: Please list any funding that was received in support of this abstract.: National Center for Advancing Translational Sciences [UL1TR000448, sub award KL2TR000450], Institute of Clinical and Translational Sciences at Washington University [UL1RR024992].