Premium
Connectomic Profiling Identifies Responders to Vagus Nerve Stimulation
Author(s) -
Mithani Karim,
Mikhail Mirriam,
Morgan Benjamin R.,
Wong Simeon,
Weil Alexander G.,
Deschenes Sylvain,
Wang Shelly,
Bernal Byron,
Guillen Magno R.,
Ochi Ayako,
Otsubo Hiroshi,
Yau Ivanna,
Lo William,
Pang Elizabeth,
Holowka Stephanie,
Snead O. Carter,
Donner Elizabeth,
Rutka James T.,
Go Cristina,
Widjaja Elysa,
Ibrahim George M.
Publication year - 2019
Publication title -
annals of neurology
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 4.764
H-Index - 296
eISSN - 1531-8249
pISSN - 0364-5134
DOI - 10.1002/ana.25574
Subject(s) - vagus nerve stimulation , receiver operating characteristic , diffusion mri , fractional anisotropy , medicine , vagus nerve , stimulation , magnetic resonance imaging , radiology
Objective Vagus nerve stimulation (VNS) is a common treatment for medically intractable epilepsy, but response rates are highly variable, with no preoperative means of identifying good candidates. This study aimed to predict VNS response using structural and functional connectomic profiling. Methods Fifty‐six children, comprising discovery (n = 38) and validation (n = 18) cohorts, were recruited from 3 separate institutions. Diffusion tensor imaging was used to identify group differences in white matter microstructure, which in turn informed beamforming of resting‐state magnetoencephalography recordings. The results were used to generate a support vector machine learning classifier, which was independently validated. This algorithm was compared to a second classifier generated using 31 clinical covariates. Results Treatment responders demonstrated greater fractional anisotropy in left thalamocortical, limbic, and association fibers, as well as greater connectivity in a functional network encompassing left thalamic, insular, and temporal nodes ( p < 0.05). The resulting classifier demonstrated 89.5% accuracy and area under the receiver operating characteristic (ROC) curve of 0.93 on 10‐fold cross‐validation. In the external validation cohort, this model demonstrated an accuracy of 83.3%, with a sensitivity of 85.7% and specificity of 75.0%. This was significantly superior to predictions using clinical covariates alone, which exhibited an area under the ROC curve of 0.57 ( p < 0.008). Interpretation This study provides the first multi‐institutional, multimodal connectomic prediction algorithm for VNS, and provides new insights into its mechanism of action. Reliable identification of VNS responders is critical to mitigate surgical risks for children who may not benefit, and to ensure cost‐effective allocation of health care resources. ANN NEUROL 2019;86:743–753