Identification of gait imagery based on fNIRS and class-dependent sparse representation

Abstract The brain-computer interface (BCI) driven by gait imagery based on functional near-infrared spectroscopy (fNIRS) has potential applications in rehabilitation training for lower limb motor dysfunction, but its performance needs to be improved. The effectiveness of class-dependent sparse representation classification (cdSRC) for identifying gait imagery was explored in the study. First, fNIRS signals were collected from 15 subjects during gait imagery (normal gait imagery and abnormal gait imagery after stroke) and idle state. Mean value, peak value and root mean square of oxyhemoglobin (HbO) and their combinations were calculated as features for classification. Class-dependent K-nearest neighbor (cdKNN) and class-dependent orthogonal matching pursuit (cdOMP) were used to solve HbO features coded by sparse representation and classify them, and the classification results were compared with those obtained by support vector machine (SVM) and KNN. The experimental results showed that the average classification accuracy of three tasks by cdSRC using the combination features of mean value, peak value and root mean square was 87.39 ± 2.59%, which was significantly higher than those achieved by SVM and KNN (78.67 ± 3.96% and 79.78 ± 4.77%, respectively). We discovered that cdSRC combined with fNIRS could effectively identify gait imagery and also proved that the combined features of HbO had better separability than a single feature for gait imagery. Recognizing gait imagery based on fNIRS can be applied to BCI to provide new control commands. This type of BCI can provide active rehabilitation training methods for the disabled, such as providing control commands for mechanical prostheses so that the disabled can perform active rehabilitation training to restore some of their motor functions. In addition, to our knowledge, this study is the first to introduce cdSRC to identify gait imagery (three classes) based on fNIRS.