Analysis methods for measuring passive auditory fNIRS responses generated by a block-design paradigm

Abstract

Significance: Functional near-infrared spectroscopy (fNIRS) is an increasingly popular tool in auditory research, but the range of analysis procedures employed across studies may complicate the interpretation of data. Aim: We aim to assess the impact of different analysis procedures on the morphology, detection, and lateralization of auditory responses in fNIRS. Specifically, we determine whether averaging or generalized linear model (GLM)-based analysis generates different experimental conclusions when applied to a block-protocol design. The impact of parameter selection of GLMs on detecting auditory-evoked responses was also quantified. Approach: 17 listeners were exposed to three commonly employed auditory stimuli: noise, speech, and silence. A block design, comprising sounds of 5 s duration and 10 to 20 s silent intervals, was employed. Results: Both analysis procedures generated similar response morphologies and amplitude estimates, and both indicated that responses to speech were significantly greater than to noise or silence. Neither approach indicated a significant effect of brain hemisphere on responses to speech. Methods to correct for systemic hemodynamic responses using short channels improved detection at the individual level. Conclusions: Consistent with theoretical considerations, simulations, and other experimental domains, GLM and averaging analyses generate the same group-level experimental conclusions. We release this dataset publicly for use in future development and optimization of algorithms.

Keywords: analysis methods; auditory responses; block-design paradigm; passive task; speech.

Fig. 1

Location of sources and detectors. Four ROI were created to cover the left IFG, the left and right STG, and the occipital lobe. Sources are shown as red dots, detectors are shown as black dots, and channels are shown as white lines with an orange dot representing the midpoint. The montage is shown from the (a) left; (b) back; and (c) right views of the brain.

Fig. 2

Summary of frequency information. The frequency content of the expected neural response based on trigger information and model HRF is shown in red (arbitrary scaling). The applied filter is shown in blue. Raw data from an example file are shown in black, with the solid line indicating the mean value across all channels and the shading representing 95% confidence intervals across channels. It is worth noting that the filter retains most of the experimental frequency content while removing high-frequency heart rate content (around 1 Hz) and low frequency content in the measured data.

Fig. 3

Morphology of auditory fNIRS responses using the averaging approach for all ROI and conditions. Each column represents a different region of interest as illustrated in the top-down head view inset. Each row represents a different stimulus condition. Red represents oxyhemoglobin, and blue represents deoxyhemoglobin. Shaded lines indicate 95% confidence intervals. Responses were observed over the left and right STG for both speech and noise conditions, but not for silence.

Fig. 4

Morphology of auditory fNIRS responses over the STG. Each column represents a different stimulus condition. Responses are illustrated for both oxy- and deoxyhemoglobin in red and blue, respectively. The shaded areas and solid line represent the mean and 95% confidence intervals for the averaging approach. The dashed lines illustrate the estimates for the FIR GLM approach. It is worth noting that the averaging and FIR GLM fits are quite similar, except for a larger estimate for the FIR approach in the speech condition.

Fig. 5

The effect of systemic response correction on auditory fNIRS response estimates. (a) ROC curves for the STG region of interest and (b) individual channels over the STG. (c) Summary statistics from the individual channel ROC with area under the curve (circle) and TPR at 5% F (square) metrics for each method. Analysis with no systemic correction is included as a reference (green), analysis with 1, 2, 4, or all PC of the short channels as regressors in the GLM is shown (orange, blue, light green, yellow, respectively), and all short channels included as individual regressors (brown) or averaged per chromophore (gray). Note that all systemic response correction approaches provide improved detection over no correction. Including all PCs, the mean of the short channels, or all individual channels provides best auditory response detection.

Fig. 6

The effect of sample rate on auditory fNIRS response estimates. (a) ROC curves for the STG region of interest and (b) individual channels over the STG. (c) Summary statistics from the individual channel ROC with area under the curve (circle) and TPR at 5% FPR (square) metrics for data sampled at different rates. Analysis indicates improved performance with increasing sample rate, but with limited improvement above $\sim 0.6\mathrm{Hz}$.

Fig. 7

The effect of boxcar function duration on auditory fNIRS response estimates. (a) ROC curves for the STG region of interest and (b) individual channels over the STG. (c) Summary statistics from the individual channel ROC with area under the curve (circle) and TPR at 5% FPR (square) metrics for different boxcar durations. Analysis indicates optimal detection rates for a 3 s boxcar function; note that the stimulus duration was 5 s.

Fig. 8

Estimates of response per condition and region of interest using the GLM analysis. Oxy- and deoxyhemoglobin responses are shown in red and blue, respectively. The presence of a response (statistical difference to zero) is indicated by a triangle. Error bars represent the 95% confidence intervals of the mean.