Synchronizability predicts effective responsive neurostimulation for epilepsy prior to treatment

Abstract

Despite the success of responsive neurostimulation (RNS) for epilepsy, clinical outcomes vary significantly and are hard to predict. The ability to forecast clinical response to RNS therapy before device implantation would improve patient selection for RNS surgery and could prevent a costly and ineffective intervention. Determining and validating biomarkers predictive of RNS response is difficult, however, due to the heterogeneity of the RNS patient population and clinical procedures; large, multi-center datasets are needed to quantify patient variability and to account for stereotypy in the treatment paradigm of any one center. Here we use a distributed, cloud-based pipeline to analyze a federated dataset of intracranial EEG recordings, collected prior to RNS surgery, from 30 patients across three major epilepsy centers. Based on recent work modelling the controllability of distributed brain networks, we hypothesize that broader brain network connectivity, beyond the seizure onset zone, can predict RNS response. We demonstrate how intracranial EEG recordings can be leveraged through network analysis to uncover biomarkers that predict response to RNS therapy. Our findings suggest that peri-ictal changes in synchronizability, a global network metric shown to accurately predict outcome from resective epilepsy surgery, can distinguish between good and poor RNS responders under the current RNS therapy guidelines (area under the receiver operating characteristic curve of 0.75). Furthermore, this study also provides a proof-of-concept roadmap for multicenter collaboration where practical considerations impede sharing datasets fully across centers.

Electrocorticography and stereo EEG provide distinct measures of brain connectivity: Implications for network models

“…Recent evidence supports the hypothesis that epilepsy arises from disordered connectivity 7,8 , and that mapping brain networks may aid in both selecting candidates for invasive treatment and identifying therapeutic targets for surgical resection, ablation or device implants 9 . In a brain network model, discrete 'nodes' exist either at the sensor-level for functional connectivity derived from iEEG signals, or at the atlas region-of-interest (ROI) level for structural connectivity derived from imaging 10 .…”