Identifying the engagement of a brain network during a targeted tDCS-fMRI experiment using a machine learning approach

by Anant Shinde, Sovesh Mohapatra, Gottfried Schlaug

Transcranial direct current stimulation (tDCS) can noninvasively modulate behavior, cognition, and physiologic brain functions depending on polarity and dose of stimulation as well as montage of electrodes. Concurrent tDCS-fMRI presents a novel way to explore the parameter space of non-invasive brain stimulation and to inform the experimenter as well as the participant if a targeted brain region or a network of spatially separate brain regions has been engaged and modulated. We compared a multi-electrode (ME) with a single electrode (SE) montage and both active conditions with a no-stimulation (NS) control condition to assess the engagement of a brain network and the ability of different electrode montages to modulate network activity. The multi-electrode montage targeted nodal regions of the right Arcuate Fasciculus Network (AFN) with anodal electrodes placed over the skull position of the posterior superior temporal/middle temporal gyrus (STG/MTG), supramarginal gyrus (SMG), posterior inferior frontal gyrus (IFG) and a return cathodal electrode over the left supraorbital region. In comparison, the single electrode montage used only one anodal electrode over a nodal brain region of the AFN, but varied the location between STG/MTG, SMG, and posterior IFG for different participants. Whole-brain rs-fMRI was obtained every three seconds. The tDCS-stimulator was turned on at 3 minutes after the scanning started. A 4D rs-fMRI data set was converted to dynamic functional connectivity (DFC) matrices using a set of ROI pairs belonging to the AFN as well as other unrelated brain networks. In this study, we evaluated the performance of five algorithms to classify the DFC matrices from the three conditions (ME, SE, NS) into three different categories. The highest accuracy of 0.92 was obtained for the classification of the ME condition using the K Nearest Neighbor (KNN) algorithm. In other words, applying the classification algorithm allowed us to identify the engagement of the AFN and the ME condition was the best montage to achieve such an engagement. The top 5 ROI pairs that made a major contribution to the classification of participant’s rs-fMRI data were identified using model performance parameters; ROI pairs were mainly located within the right AFN. This proof-of-concept study using a classification algorithm approach can be expanded to create a near real-time feedback system at a participant level to detect the engagement and modulation of a brain network that spans multiple brain lobes.