Investigation of functional near-infrared spectroscopy signal quality and development of the hemodynamic phase correlation signal

Abstract

Significance: There is a longstanding recommendation within the field of fNIRS to use oxygenated ( ${\mathrm{HbO}}_2$ ) and deoxygenated (HHb) hemoglobin when analyzing and interpreting results. Despite this, many fNIRS studies do focus on ${\mathrm{HbO}}_2$ only. Previous work has shown that ${\mathrm{HbO}}_2$ on its own is susceptible to systemic interference and results may mostly reflect that rather than functional activation. Studies using both ${\mathrm{HbO}}_2$ and HHb to draw their conclusions do so with varying methods and can lead to discrepancies between studies. The combination of ${\mathrm{HbO}}_2$ and HHb has been recommended as a method to utilize both signals in analysis. Aim: We present the development of the hemodynamic phase correlation (HPC) signal to combine ${\mathrm{HbO}}_2$ and HHb as recommended to utilize both signals in the analysis. We use synthetic and experimental data to evaluate how the HPC and current signals used for fNIRS analysis compare. Approach: About 18 synthetic datasets were formed using resting-state fNIRS data acquired from 16 channels over the frontal lobe. To simulate fNIRS data for a block-design task, we superimposed a synthetic task-related hemodynamic response to the resting state data. This data was used to develop an HPC-general linear model (GLM) framework. Experiments were conducted to investigate the performance of each signal at different SNR and to investigate the effect of false positives on the data. Performance was based on each signal's mean $T$ -value across channels. Experimental data recorded from 128 participants across 134 channels during a finger-tapping task were used to investigate the performance of multiple signals [ ${\mathrm{HbO}}_2$ , HHb, HbT, HbD, correlation-based signal improvement (CBSI), and HPC] on real data. Signal performance was evaluated on its ability to localize activation to a specific region of interest. Results: Results from varying the SNR show that the HPC signal has the highest performance for high SNRs. The CBSI performed the best for medium-low SNR. The next analysis evaluated how false positives affect the signals. The analyses evaluating the effect of false positives showed that the HPC and CBSI signals reflect the effect of false positives on ${\mathrm{HbO}}_2$ and HHb. The analysis of real experimental data revealed that the HPC and HHb signals provide localization to the primary motor cortex with the highest accuracy. Conclusions: We developed a new hemodynamic signal (HPC) with the potential to overcome the current limitations of using ${\mathrm{HbO}}_2$ and HHb separately. Our results suggest that the HPC signal provides comparable accuracy to HHb to localize functional activation while at the same time being more robust against false positives.

Keywords: functional near-infrared spectroscopy; hemodynamic response; neuroscience.

Fig. 1

Percentage of studies using each signal. Peer-reviewed studies from 2017 to 2020 in the field of cognitive neuroscience using fNIRS were reviewed. Resting-state papers were excluded.

Fig. 2

Synthetic data formation. Resting-state fNIRS data were obtained from the Hitachi WOT system using 16 channels at 5-Hz sampling frequency.

Fig. 3

Generation of HPC signal from its component signals, (a) ${\mathrm{HbO}}_2$ and (b) HHb. A task frequency of 0.025 Hz was used. Denoted by the shaded areas in the time series and highlighted by a white box in the spectrograms. Continuous wavelet transform applied to both signals (c) and (d). Cross wavelet transform is computed (e) and the phase extracted is shown in (f), which is converted to a normalized value, the HPC, in (g).

Fig. 4

Pipeline used to determine which frequency should be used as the measured signal in the GLM. HPC was computed for all 16 channels of all 18 datasets. The GLM was used to output $\beta$ values for all frequencies of all HPC transforms. The index of the maximum $\beta$ value was determined. The index corresponds to the frequency at which that $\beta$ value was found. The most commonly occurring index was found for each dataset, across channels. The task periods are highlighted in gray in the time data. The outlined area in the HPC spectrograms correspond to frequencies of 0.0234, 0.0250, and 0.0268 Hz. All three are outlined to account for the indirect scale to frequency conversion. The corresponding task-frequency is highlighted in yellow for the HPC spectrograms.

Fig. 5

Full workflow of analysis using HPC signal. (a) The HPC is computed using the preprocessed measured ${\mathrm{HbO}}_2$ and HHb signals for a given channel. The relevant task frequency is identified, and the HPC signal according to that frequency (or scale) is extracted from the full spectrogram. (b) To compute the HPC-RF the standard HRF of ${\mathrm{HbO}}_2$ and HHb are first computed using the task timings, the HPC transform is then applied using the standard HRFs as inputs. (B.i) The reversal artifact detection function can be used at this stage to evaluate blocks that may contain the reversed artifact. Users have the option to automatically remove these blocks from the HPC transform, or to visually inspect blocks using the output as a guide and then remove from analysis. (c) The standard GLM is used with the Measured HPC at a task frequency as the measured parameter ($Y$ in GLM) and the HPC-RF as the input to the design matrix ($X$ in GLM). Once the $\beta$ values are computed standard statistical analysis can be carried out, and a threshold $\beta$ value can be applied, the authors suggest a value of ${\beta }_{\text{Thresh}}\ge 0.3$.

Fig. 6

Example HPC output. The plots on the left show the resting state fNIRS data (see Fig. 1) and the corresponding HPC signal. The right shows the same data after the resting state fNIRS data has been added to the synthetic task HRFs with a task frequency of 0.025 Hz. In both HPC, the task frequency is highlighted by a yellow box and zoomed in below. The time data also show the task blocks shaded in gray.

Fig. 7

(a) Frequency check results. (b) Activation threshold results; the points correspond to the means as explained in Fig. 4, and the error-bars represent standard error of the mean. The red line in (a) displays the expected task related frequency (0.025 Hz), while (b) displays the chosen threshold based on the results.

Fig. 8

Percentage deviation of all signals when varying the amplitude of the task-related component of the synthetic data. Amplitude 1 has the highest signal-noise ratio. $T$-values were computed using a one-sample two-tailed $t$-test, with a hypothesized mean of 0.

Fig. 9

(a) Results from reversing the sign of the HHb HRF and (b) results from reversing the sign of the ${\mathrm{HbO}}_2$ signal. The percentages listed in the legend refer to what percentage of the time series was convolved with a task-related HRF, 90% corresponds to near full activation. $T$-values were computed using a one-sample two-tailed $t$-test with hypothesized mean equal to 0.

Fig. 10

All signals, FDR corrected $p\text{-}\text{value}\le 0.01$. (a) ${\mathrm{HbO}}_2$, (b) HHb, (c) ${\mathrm{Hb}}_{\mathrm{Diff}}$, (d) ${\mathrm{Hb}}_{\text{Total}}$, (e) CBSI, and (f) HPC. The experimental task was a right-handed finger-thumb tapping task. Expected channel activations were in the left primary motor cortex.