Multi-class Classification of Motor Execution Tasks using fNIRS

This paper investigates the problem of classification of multi-class movement execution tasks from signals obtained via functional near infrared spectroscopy (fNIRS). fNIRS data is acquired from five healthy subjects while performing four types of motor execution tasks as well as a non-movement task (five classes in total). Various feature sets are extracted based on the mean of changes in the concentration of oxygenated hemoglobin ([ΔHbO]) signals computed across the [0 - 2], [1 - 3], and [2 - 4] sec intervals. A multi-class support vector machine classifier with a quadratic polynomial kernel (QSVM) is utilized to classify movement and non-movement classes (total of 5 classes) using the data from the three time intervals. Classification results revealed that the average accuracy obtained for data using [2 - 4] sec interval is higher than the other two (78.55%). In addition, a comparison between the classification results of the data obtained from only the motor cortex vs from multiple regions of the brain is done. Our results demonstrate that by using fNIRS data from different regions of the brain, the classification accuracy is improved by 10 - 12% as compared to the case when the data is used only from the motor region.