Multivariate Kalman filter regression of confounding physiological signals for real-time classification of fNIRS data

SIGNIFICANCE: Functional near-infrared spectroscopy (fNIRS) is a noninvasive technique for measuring hemodynamic changes in the human cortex related to neural function. Due to its potential for miniaturization and relatively low cost, fNIRS has been proposed for applications, such as brain–computer interfaces (BCIs). The relatively large magnitude of the signals produced by the extracerebral physiology compared with the ones produced by evoked neural activity makes real-time fNIRS signal interpretation challenging. Regression techniques incorporating physiologically relevant auxiliary signals such as short separation channels are typically used to separate the cerebral hemodynamic response from the confounding components in the signal. However, the coupling of the extra-cerebral signals is often noninstantaneous, and it is necessary to find the proper delay to optimize nuisance removal. AIM: We propose an implementation of the Kalman filter with time-embedded canonical correlation analysis for the real-time regression of fNIRS signals with multivariate nuisance regressors that take multiple delays into consideration. APPROACH: We tested our proposed method on a previously acquired finger tapping dataset with the purpose of classifying the neural responses as left or right. RESULTS: We demonstrate computationally efficient real-time processing of 24-channel fNIRS data (400 samples per second per channel) with a two order of selective magnitude decrease in cardiac signal power and up to sixfold increase in the contrast-to-noise ratio compared with the nonregressed signals. CONCLUSION: The method provides a way to obtain better distinction of brain from non-brain signals in real time for BCI application with fNIRS.