Online classification of the near-infrared spectroscopy fast optical signal for brain-computer interfaces

Abstract

Objective. The fast optical signal (FOS), measured with near-infrared spectroscopy (NIRS), has high temporal and competitive spatial resolution which provides an opportunity for a novel braincomputer interface modality. However, the reliability of the FOS has been debated due to its low signal-to-noise ratio. Approach. This study examined the feasibility of automatically classifying the prefrontal FOS response during a visual oddball task. FOS measurements were collected from 15 participants during 3 offline and 2 online sessions. Classification feedback was provided to participants during online sessions. The FOS classification algorithm discriminated between oddball and frequent responses. DC intensity and phase delay FOS measurements were classified individually with both support vector machine (SVM) and linear discriminant analysis (LDA) classifiers. The decisions of these classifiers were combined with a weighted majority vote. Fifteen-trial averages were selected for optimal classification results and the best feature types were found to be the number of zero crossing and the variance. Main results. FOS responses to oddball and frequent images were classified offline with an average balanced accuracy of 62±5% and classified online with an averaged balanced accuracy of 63±6% across all participants. Offline classification accuracies were significantly higher than chance for all participants. Online classification results were significantly above chance in both online sessions for 7 of 14 participants. Event-related potentials (ERPs) were classified using a similar algorithm with an average balanced accuracy of 77±5%, confirming that the prefrontal neuronal response to the visual oddball task could be classified above the level required (>70%) for effective BCI communication. Significance. The FOS classification results demonstrated that automatic classification of the FOS is possible at above-chance levels, however, FOS classification accuracies did not reach the effective BCI communication threshold. Further FOS classification efforts should focus on investigating spectral features as well as adding measurement channels over the fronto-central and parietal areas of the brain.

Assessing Time-Resolved fNIRS for Brain-Computer Interface Applications of Mental Communication

“…Both algorithms produced similar accuracies (76% for SVM and 75% for LDA); however, SVM resulted in a better balance between sensitivity and specificity (79% and 71%, respectively) compared to LDA (83% and 58%, respectively). Overall, these estimates of classification accuracy are in-line with previous reports (Naseer and Hong, 2015a) and meet the minimum threshold of 70% for a BCI to be considered effective for communication (Proulx et al, 2018). Although the classification accuracy is comparable to results from other fNIRS studies involving various activation tasks for mental communication (Naseer and Hong, 2015b), it was less than the accuracy reported in an fMRI study involving the same tennis imagery task (Monti et al, 2010).…”

Methodologies on the Enhanced Spatial Resolution of Non-Invasive Optical Brain Imaging: A Review

“…This is probably a major advantage as FOI may render an ideal foundation technology for brain imaging data fusion [228]. With the help of this methodology, multiple researchers [204], [225], [229]- [231] have shown that fast optical signals can be measured consistently with combined high spatial and temporal resolution. However, some research groups have also been questioned these possibilities [232], [233].…”

Fast Optical Signals for Real-Time Retinotopy and Brain Computer Interface

“…Among the variety of classifiers available, the support vector machine (SVM) has been demonstrated to be highly effective for classifying brain signals [22][23][24][25][26][27]. Proulx et al (2018), for example, used SVM and Linear Discriminant Analysis (LDA) using FOS (light intensity and phase delay, a measure of average photons time of flight using frequency-domain NIR systems) to distinguish between unusual and common responses [28]. The outcomes were combined using a weighted majority vote.…”

“…The outcomes were combined using a weighted majority vote. FOS responses to rare and common images were classified among participants either offline, obtaining an average balanced accuracy of 62.5%, or online, with an average balanced accuracy of 63.6% [28]. The current study aimed to investigate the capability of combining FOS with SVM for single-trial retinotopy estimation in BCI applications.…”