Resting-State-Based Spatial Filtering for an fNIRS-Based Motor Imagery Brain-Computer Interface

Functional near-infrared spectroscopy (fNIRS) has attracted much attention in brain-computer interface (BCI) area due to its advantages of portability, robustness to electrical artifacts, etc. However, in practical applications, fNIRS-based BCI usually needs a labor-intensive and time-consuming training session (calibration procedure) to optimize the user-specific neural spatial and temporal patterns for further classification. Recently, studies revealed that neural spatial and temporal patterns extracted from user-specific resting-state brain signals were closely related to those of his/her task data. In this study, we proposed a resting-state independent component analysis (RSICA) based spatial filtering algorithm aiming at extracting individual task-related spatial and temporal brain patterns from the resting-state data. Specifically, independent component analysis (ICA) was applied to extract different independent components (ICs) from resting-state fNIRS data. The ICs with their spatial filter weights maximally lateralized over the sensorimotor regions were regarded as most relevant to motor imagery. These spatial filters were used to spatially filter the multi-channel motor imagery task data for feature extraction. Based on 8-minute resting-state data and a small training dataset (20 trials) from 10 participants, the proposed RSICA algorithm achieved an approximately 7% increase in left vs. right hand motor imagery classification accuracy, as compared to the conventional common spatial pattern (CSP)-based and shrinkage algorithms (69.8±12.1%, 63.3±10.3% and 63.4±11.8%, respectively). For acquiring a similar level of classification accuracy (i.e. 70%), the number of training data required could be reduced from 36 trials (CSP) to 22 trials (RSICA). As a relatively small training set is required to obtain a satisfactory performance, training burden is significantly reduced by RSICA, which might be useful for developing practical fNIRS-based motor imagery BCIs.