A pilot study for active muscles decoding using functional near-infrared spectroscopy

Abstract

This study is a preliminary step toward gait identification using a non-invasive brain-computer interface. We investigated the feasibility of decoding different active muscles from brain activation using functional near-infrared spectroscopy (fNIRS). A two-section experiment was designed to alternately activate the subjects' hamstring and quadriceps. Nine right-handed subjects, aged $28.1\pm 3.5$, were recruited for the experiment. The measured optical intensities were converted to optical density changes and filtered by targeted principal component analysis (tPCA), a lowpass filter, and a highpass filter sequentially. Six features (slope, skewness, kurtosis, peak-to-peak, standard deviation, and entropy) were extracted from the filtered signals and fed to a linear discriminant analysis (LDA) classifier in pairs. Results showed that using the feature pair of slope-standard deviation, we could achieve a classification rate of more than 80% for all four categories (sitting extension, sitting flexion, standing extension, and standing flexion). The maximum classification accuracy was 85.34% for training validation and 92.22% for the testing dataset. Subsequently, an ANOVA test found significant decoding differences among feature combinations. Additionally, no significant difference is found among slope-included feature pairs, skewness-standard deviation, and standard deviation-entropy. The results proved that decoding different muscles related to gait is possible using fNIRS in the future.