Authors :
Presenting Author: Qirui Zhang, MD – Thomas Jefferson University
Sam Javidi, PhD – Farber Institute for Neuroscience, Department of Neurology – Thomas Jefferson University; Ankeeta A., PhD – Farber Institute for Neuroscience, Department of Neurology – Thomas Jefferson University; Xiaosong He, PhD – Department of Psychology – University of Science and Technology of China; Yolanda Kry, BA – Farber Institute for Neuroscience, Department of Neurology – Thomas Jefferson University; Michael Sperling, MD – Department of Neurology – Thomas Jefferson University; Joseph Tracy, PhD – Farber Institute for Neuroscience, Department of Neurology – Thomas Jefferson University
Rationale:
Medically intractable temporal lobe epilepsy (TLE) is commonly treated through surgery, such as anterior temporal lobectomy (ATL). However, the postsurgical seizure freedom rate has remained stable at 48-76% for decades, despite improvements in diagnostic imaging and intracranial monitoring. To achieve seizure freedom it is likely that the brain regions not touched by surgery will have to alter their functional properties, changing, for instance, inter-regional communication to now identify/combat epileptogenic signaling. We hypothesize that the underlying capacity/propensity to undergo such functional changes postsurgery may be detectable prior to surgery using measures sensitive to brain network flexibility and dynamics. Here, we measure such network dynamics and test for their ability to predict surgical outcomes. Our goal is to improve prognostication by decoding prior to surgery the network phenotypes that implement the brain resiliency and change necessary to produce good outcomes.
Methods:
Sample was comprised of 67 ATL treated TLE patients (right=34; left=33) and matched healthy controls (HC, n= 32) with presurgical resting state fMRI (rs-fMRI) data. We calculated integration and recruitment of dynamic community structure on rs-fMRI data (232 brain regions assigned to 18 rs-fMRI cortical and sub-cortical networks). Using the W-score we calculated the deviation of patients' metrics relative to HC, controlling for sex and age. Using machine learning (XGboost, a decision-tree-based ensemble algorithm; 5-fold cross-validation) we compared the single and combined predictive capacity of two sets of data, clinical/epilepsy characteristics and our dynamic rs-fMRI network features. Post-operative seizure status at 13-18 months was our response variable.
Results:
Seizure outcome prediction using only clinical characteristics was mediocre (AUC = 0.64). In contrast, the predictive capacity of the dynamic rs-fMRI metrics was more robust (AUC = 0.89). By taking the top 10 predictive features from each single set model, we formed a combined model with excellent predictive power (AUC = 0.95). The Delong test showed that the rs-fMRI and the combined model had significantly higher AUC compared to the clinical model alone (p < 0.001). Specific integrations such as sub-cortical to Limb_B (temporal pole) displayed the highest feature importance, suggesting presurgical indexing of these connectivities may be of critical importance to prognostication. Several rs-fMRI metrics were significantly associated with clinical features (p < 0.05).