Abstract:
Functional near-infrared spectroscopy (fNIRS) is a portable non-invasive modality used excessively in modern brain computer interface (BCI) systems with ultimate goal of aiding people suffering from temporary or permanent disabilities. The brain data acquired using fNIRS devices is associated with the user intentions through mental activation and then, decoded into control commands. Classification accuracy is one of the measures of state-of-the-art BCI systems, identifying region of interest (ROI) or channel of interest (COI) and optimal channel selection may enhance classification performance. Effectiveness of fNIRS-BCI can be substantially increased by selecting cortical activity based channel selection methods. This research is aimed at selection of optimal channels for fNIRS-BCI system by proposing z-score method to improve classification accuracy. In the proposed method, cross correlation technique is applied between acquired brain signal and desired hemodynamic response function (dHRF). The z-scores of maximum values of correlation coefficients for each channel are calculated and channels showing positive z-scores are selected for further processing. Comparative analysis of the proposed method is done with already existing t-value method and it is further validated by calculating classification accuracies without channel selection technique i.e. using all channels. The study utilizes open access dataset that contained brain signals of 17 healthy subjects for a two-class fNIRS-BCI problem (task vs rest); right-hand finger tapping (RFT), left-hand finger tapping (LFT) and foot tapping (FT). Conventional statistical features i.e. mean, peak, slope, skewness, kurtosis and variance of the oxygenated hemoglobin (HbO) signal were used as features for classification by linear discriminant analysis (LDA). The classification accuracies achieved by using the proposed method for RFT vs rest, LFT vs rest, and FT vs rest tasks were 72.24 ± 6.2%, 72 ± 8.1% and 69.41 ± 6.7%, respectively. The results yielded by the z-score method have shown a considerable improvement in classification performance as a step forward for performance enhancement for BCI systems.
