Work

A deep learning model for a priori estimation of spatiotemporal regions for neuroimaging guided non-invasive brain stimulation

Year: 2021

Type: article

Abstract: Background: In-vivo effects of transcranial electrical stimulation do not change firing rates[1] but modulate endogenous brain activity[2]. Therefore, brain-state-dependent non-invasive brain stimulation(NIBS) may be guided by simultaneous neuroimaging; however, the challenge remains in capturing nonlinear spatiotemporal interactions[3]. We present an explainable deep learning model for a-priori estimation of behavior-related spatiotemporal brain targets. Methods: Explainable deep learning(DL) model simultaneously predicts the behavior from the brain activation measurements and identifies the segment of the signal that contributes the most to the prediction. The model employs an attention-based Long Short-Term Memory(LSTM) recurrent autoencoder to encode the context of multivariate brain imaging signals for behavior predictions. A novel aspect of this model is the attention mechanism that not only enhances the generalizability of behavior prediction but also identifies the segments of the signals that are most significant for behavior prediction. The model is developed using a dataset comprised of functional Near-Infrared Spectroscopy(fNIRS) measurements from the prefrontal cortex(PFC) region from eight medical students performing pattern cutting Fundamentals of Laparoscopic Surgery(FLS) task in a non-stimulated environment. Performance scores, which encapsulate the behavior of pattern cutting tasks, are calculated using a proprietary scoring rubric that was obtained from the FLS committee under a non-disclosure agreement. Results: The multivariate analysis using PFC channels has revealed that early PFC activation(<10s) contributes the most to the prediction of task performance (R2=0.87; pass-fail classification accuracy=99%). Discussion: The identification of the sequence of the signal that contributes the most in mapping brain activation to the behavior will be instrumental in delivering NIBS. Here, the initial skill learning stage with primarily PFC activation is postulated to require DL neuroimaging-guided NIBS of the functional PFC subdomains [4]. References: [1]Krause et al.,doi: 10.1016/j.cub.2017.09.020. [2]Bikson, et al.,doi: 10.1007/978-3-319-95948-1_3. [3]Dutta,doi: 10.3389/fnsys.2015.00107. [4]Koechlin and Summerfield,doi: 10.1016/j.tics.2007.04.005. Keywords: explainable deep learning model , functional Near-Infrared Spectroscopy, personalized NIBS, attention based LSTM

Source: Brain stimulation

DOI: 10.1016/j.brs.2021.10.323