Multiclass classification of hemodynamic responses for performance improvement of functional near-infrared spectroscopy-based brain-computer interface

We improved the performance of a functional near-infrared spectroscopy (fNIRS)-based brain-computer interface based on relatively short task duration and multiclass classification. A custom-built eight-channel fNIRS system was used over the motor cortex areas in both hemispheres to measure the hemodynamic responses evoked by four different motor tasks (overt execution of arm lifting and knee extension for both sides) instead of finger tapping. The hemodynamic responses were classified using the naive Bayes classifier. Among the mean, max, slope, variance, and median of the signal amplitude and the time lag of the signal, several signal features are chosen to obtain highest classification accuracy. Ten runs of threefold cross-validation were conducted, which yielded classification accuracies of 87.1% ± 2.4% to 95.5% ± 2.4%, 77.5% ± 1.9% to 92.4% ± 3.2%, and 73.8% ± 3.5% to 91.5% ± 1.4% for the binary, ternary, and quaternary classifications, respectively. Eight seconds of task duration for obtaining sufficient quaternary classification accuracy was suggested. The bit transfer rate per minute (BPM) based on the quaternary classification accuracy was investigated. A BPM can be achieved from 2.81 to 5.40 bits/min.