Source Distribution Association Algorithm Based on Independent Component Analysis

To realize the transfer between EEG and fNIRS signals, we first need to adjust the positions and numbers of the two signals’ channels to be consistent. Although EEG and fNIRS channels show one-to-many correspondence, which means that each EEG channel is associated with multiple fNIRS channels. But for the convenience of calculation, we need to make a one-to-one correlation. Therefore, this paper proposes channel matching from the perspective of source distribution. This section will introduce the channel adjustment strategy in detail. ICA is widely used as a popular technique to remove artifacts in the BCI field. In addition, for the motor imagery task, because the interstitial activity is significant and its source can usually be considered specific, the ICA algorithm is also suitable for extracting the exciting regions. ICA is a generative model, which describes how to generate observation data by mixing independent components (Hyvarinen and Oja, 2000). Considering the calculation speed, the ICA algorithm used in this article is FastICA (Dominic et al., 2010). In our study, we first obtain the source signal through the ICA algorithm. We believe that both EEG and fNIRS are based on motor imagery tasks, and the brain regions where the source signals generated have some certain consistency. We then matched channels by correlation of source signals.

As described in Algorithm 1, we first used the ICA algorithm on the EEG signals to get eight independent sources (the number of channels is eight), which were marked as IC₁∼IC₈ in turn. Then, draw the power spectral density of these independent components. As shown in Figure 4, the power spectra of different independent components vary greatly. Calculate the distance between different types of power spectral density (PSD) curves on the power spectrum (the distance is calculated as the mean distance across all the frequencies). Then select the most obvious components that distinguish the left and right hands, and treat them as the most relevant source of classification, which is marked as S_eeg1, S_eeg2.

Source Distribution Association Algorithm Based on ICA.

The power spectra of the independent components.

As shown in Figure 5, in the fNIRS data, using the left and right-handed tasks did not work because the response of the fNIRS signal was slower than that of the EEG signal. For the fNIRS signals, ICA decomposition was performed first and then calculated the change of the source signal’s oxyhemoglobin concentration between the task period and rest period. Like EEG, the two components with the largest variation were picked up as the most relevant sources of classification. Denoted these two components as S_fnirs1, S_fnirs2.

The power spectrum comparison of left and right handed and task-resting in fNIRS data.

Usually, when we compare the columns of the mixing matrix A in ICA, we can analyze how much the source data contributes to each channel of the observation data. Therefore, the columns of the mixing matrix can be used to calculate and visualize the topography of components. Motivated by the theory above, we observed the columns in the mixing matrix belong to the four source signals and sorted the channels according to their weights. For example, the pattern diagram of S_eeg1 is shown in Figure 6. We sorted the channels according to the weight, the order of the channels is “FCC5h,” “FCC6h,” “CCP4h,” “CCP5h,” “FCC3h,” “FCC4h,” “CCP6h,” “CCP3h.” We deduced the channel sequence corresponding to the other three source signals by analogy. By combining them, the purpose of matching the EEG optode and the fNIRS channel was achieved.

The pattern diagram of Seeg1.