Predicting the epileptic seizure onset zone with brain-wide alterations of temporal dynamics in fMRI

2319 

Abstract Submission 

Karl-Heinz Nenning¹, Erkam Zengin², Ting Xu³, Gelana Tostaeva⁴, Elisabeth Freund⁴, Stanley Colcombe¹, Ashesh Mehta⁴, Michael Milham³, Stephan Bickel⁴

¹Nathan Kline Institute, Orangeburg, NY, ²Northwell Health, Manhasset, NY, ³Child Mind Institute, New York, NY, ⁴The Feinstein Institutes for Medical Research, Manhasset, NY

Karl-Heinz Nenning  
Nathan Kline Institute
Orangeburg, NY

Erkam Zengin  
Northwell Health
Manhasset, NY

Ting Xu  
Child Mind Institute
New York, NY

Gelana Tostaeva  
The Feinstein Institutes for Medical Research
Manhasset, NY

Elisabeth Freund  
The Feinstein Institutes for Medical Research
Manhasset, NY

Stanley Colcombe  
Nathan Kline Institute
Orangeburg, NY

Ashesh Mehta  
The Feinstein Institutes for Medical Research
Manhasset, NY

Michael Milham  
Child Mind Institute
New York, NY

Stephan Bickel  
The Feinstein Institutes for Medical Research
Manhasset, NY

Epilepsy has been recognized as a network disease that can disturb brain regions beyond a focal seizure onset. Previous studies linked an altered autocorrelation function (ACF) of brain activation to disturbed brain dynamics in the seizure onset region (Nedic et al, 2015) and beyond (Xie et al, 2023). Here, we used preoperative resting-state functional magnetic resonance imaging (rs-fMRI) to quantify brain-wide ACF decay rates in medically refractory epilepsy patients with medial temporal lobe seizure onset. We evaluated how brain dynamics may be disrupted due to the underlying disease, and determined the potential use of ACF decay rates to identify seizure onset zones (SOZ) to inform intervention strategies.

We studied rs-fMRI data from 15 patients with unilateral mesial temporal lobe epilepsy (TLE; 10 left) that was confirmed by intracranial stereo EEG. For each voxel, we established a feature vector characterizing the temporal dynamics based on different ACF decay rate measures (Watanabe et al, 2019; Raut et al, 2020; Ito et al, 2020). We utilized data from a group of 652 healthy controls (Cam-CAN; Shafto et al, 2014; Taylor et al, 2017) as a normative baseline, and, for each patient and voxel, quantified the deviation of the ACF decay rates as a timescale anomaly score (z-score based on the normative distribution). For each intracranial EEG electrode, we calculated the corresponding fMRI timescale anomalies and, using logistic regression, evaluated their predictive performance to classify electrodes that map the potential SOZ. Finally, we associated brain-wide timescale anomaly maps with outcome measures to examine the potential added value of preoperative rs-fMRI to guide neurosurgical intervention.

Overall, we observed reduced ACF decay rates for electrodes that are located in brain regions associated with a seizure onset zone, suggesting a more constrained temporal dynamic. Brain regions that map to SOZ-related electrodes also show a reduced regional homogeneity, emphasizing disturbances in brain activity and functional connectivity. Importantly, in a leave-one-patient-out framework, we found that ACF decay rate measures were sensitive to focal alterations and predicted the electrodes identified as seizure onset well. In 13 out of the 15 patients, we observed a prediction performance of AUC > 0.7 and that was better than chance (prediction based on shuffled labels).

Our preliminary findings revealed widespread alterations of neural dynamics in patients with temporal lobe epilepsy. Brain regions in the seizure onset zone showed a more constrained activity (slower temporal autocorrelation decay rates) and lower regional homogeneity than regions located outside the seizure onset zone. These preliminary results indicate that alterations of temporal dynamics show promise for non-invasively delineating seizure onset zones from preoperative rs-fMRI. The observed alterations also emphasize the notion of epilepsy as a network disease, affecting brain regions beyond an obvious focal seizure onset. False positive classification remains a challenge that can likely be informed with patient outcomes to determine potential electrode mislabeling and latent seizure onsets.

Classification and Predictive Modeling ²

BOLD fMRI ¹

Epilepsy