Award: American Academy of Pediatrics Award

Authors: Karim Mithani; Mirriam Mikhail, BSc; Benjamin Morgan; Simeon Wong; Alexander Weil, MD; Shelly Wang, MD, MPH; Byron Bernal, MD; Magno Guillen, PhD; James Rutka, MD, PhD; Cristina Go, MD; Elysa Widjaja, MD, MPH; George Ibrahim, MD, PhD (Toronto, Canada)

Introduction: Response to vagus nerve stimulation (VNS) in children with epilepsy is heterogeneous and its mechanism of action remains unclear. Here, we test the hypothesis that intrinsic differences in brain structural connectivity are associated with VNS response and may be used to predict outcomes in individual children. Methods: We analyzed pre-operative diffusion tensor imaging (DTI) scans from 38 children who underwent VNS for intractable epilepsy. Voxelwise statistics were generated using tract-based spatial statistics. Significantly different tracts ( p <0.05) were used to generate a support vector machine (SVM) classifier to predict treatment response. A separate cohort of 16 patients from three separate institutions were used to validate this model. We also compared patients’ DTI scans to 38 age-matched controls using the same voxelwise methodology. Finally, we conducted a principal component analysis of 31 clinical characteristics, and generated another SVM using significant components (eigenvalue >1) for comparison. Results: VNS responders showed significantly greater fractional anisotropy in 10 white matter bundles, including in limbic, thalamocortical, and hemispheric association circuitry. The resulting SVM revealed 89.5% accuracy, area under the ROC curve 0.92, 96% sensitivity, and 73% specificity on 10-fold cross-validation. When tested against 16 patients from different institutions, this classifier correctly predicted the outcome of 13 individuals (81.3% accuracy). In contrast, the SVM of clinical data showed an area under the ROC curve of 0.47 on cross-validation. Notably, whole-brain structural connectivity patterns of VNS responders more closely resembled healthy controls than non-responders did. Conclusion: These findings further our understanding of the mechanism of action of VNS, and provide a method for predicting treatment responsiveness. This is an improvement over clinical characteristics alone, which have little predictive value. Identification of accessible biomarkers to predict response to VNS is imperative to ensure that children are not exposed to unnecessary risks, and that health systems are operating efficiently.