Functional Near-Infrared Spectroscopy (fNIRS) is an optical neuroimaging technology that has rapidly gained momentum within the last decades (Boas et al., 2014; Scholkmann et al., 2014; Yücel et al., 2017). It is a non-hazardous and non-invasive optical brain imaging technique that uses near-infrared light to measure local cortical concentration changes of oxygenated and deoxygenated hemoglobin (HbO₂/HbR), which are associated with brain metabolism (Villringer and Chance, 1997; Ferrari and Quaresima, 2012). fNIRS has been considered a cost-effective and mobile alternative for functional Magnetic Resonance Imaging in conventional neuroscientific research. It is very suitable—and thus increasingly being used—for single trial analysis and Brain Computer Interface (BCI) applications (Matthews et al., 2008; Hong et al., 2018) as a single modality or along with Electroencephalography (EEG). While EEG and fNIRS signal processing is essential to increase the contrast to noise ratio (CNR) of measured brain responses, the nature of the signals and processing methods differ greatly. Hemodynamic brain responses in fNIRS are usually masked by local and systemic physiological confounding signals, for instance from superficial (scalp) blood flow, low frequency oscillations (Mayer waves), motion and breathing (Elwell et al., 1999; Yücel et al., 2016; von Lühmann et al., 2019). New and increasingly complex and powerful statistical methods are being developed that aim to remove the confounding factors in the signal, improve CNR and increase the detection/classification accuracy of hemodynamic responses. An objective way of validating the power of these novel methods and comparing them with the existing ones is to use an fNIRS dataset which has all the confounding signals but also a known hemodynamic brain response. One solution for this problem is to generate realistic fNIRS ground truth data by modeling a hemodynamic response function (HRF) on top of real resting state data (Gagnon et al., 2012; von Lühmann et al., 2019, 2020a,b). This approach can be used as a good approximation for realistic fNIRS signals with evoked responses, for which the ground truth is available. Generating such data is comparatively straight forward but requires prior knowledge in fNIRS signal characteristics as well as experience in fNIRS signal processing. Moreover, the use of short-separation fNIRS measurements and additional physiological signals, such as accelerometer or photoplethysmography (PPG), has been shown to enable methods that yield improved CNR (Yücel et al., 2015; von Lühmann et al., 2020a), but there are only few openly available multimodal fNIRS datasets (Shin et al., 2017, 2018) and even fewer multimodal datasets that include sufficient resting state periods to enable the approach described above. Thus, as a remedy, here we provide such a multimodal dataset with (and without) added synthetic HRF ground truth, short-separation fNIRS measurements, accelerometer, and other physiological measurement, for the data science community in order to facilitate the validation of novel methods. We also provide a simple code example to enable customization and modification of the HRF ground truth in the data.

The resting state data consists of two subsets: Dataset I, with 5 min resting state data from 14 participants and Dataset II, with 10 min resting state data from 14 participants. The data details follow and are summarized in Figure 1.

Baseline analysis and data quality assessment was performed using the openly available Homer3 toolbox (https://github.com/BUNPC/Homer3) (Huppert et al., 2009). HRFs were recovered from the augmented resting state data using the processing stream provided in the repository (tCCA_xmpl_procStream_Gauss_noHPF.cfg under “code” folder). This processing stream includes 0.5 Hz zero phase low pass filter with an effective order of 6, conversion to HbO₂ and HbR using the modified Beer-Lambert Law (Delpy et al., 1988; Boas et al., 2004), and subsequent HRF estimation with the temporally embedded General Linear Model (tCCA GLM) approach using short-distance channels and a polynomial drift term for nuisance regression and Gaussian basis functions for the HRF regressor (von Lühmann et al., 2020a). Figure 2A exemplifies the resulting HRF estimates in one augmented channel for all three amplitudes from participant 33 from Dataset I and participant 98 from Dataset II. Data quality is provided for each participant as across-channel average of the Signal to Noise Ratio (SNR) in Figure 2B and as the mean motion ratio across channels in Figure 2C. Channel SNR is calculated as 20 × log10 of the mean over std. of the raw intensity signal. The motion ratio is calculated as the ratio between the cumulative time of segments in the data that were considered to be confounded by motion artifacts, as identified by the Homer2 function hmrMotionArtifactByChannel (with tMotion = 0.5, tMask = 0.5, STDEVthresh = 20, AMPthresh = 5), to the total acquisition time. Figure 2D displays a typical segment of all available signals (z-scored) in the first participant in Dataset I. Long and short-separation fNIRS channels exhibit typical low frequency components and cardiac pulsation, which is also present in the PPG and BP measurement.

We reported a multimodal fNIRS resting state dataset from 28 participants, that we provide with and without added synthetic HRF ground truth at three different amplitudes. We include the script used for the generation of these data to enable users to adapt this approach to their own needs. The availability of multiple auxiliary biosignals, such as motion (accelerometer) and PPG in the data, can be used to explore and extend existing multimodal fNIRS-based signal processing approaches (von Lühmann et al., 2019, 2020a). Resting fNIRS data with added known HRF enables the validation of novel processing methods for single trial HRF detection and BCI as well as more general artifact rejection and preprocessing approaches and their comparison with existing methods. This can also be useful for methods that tackle challenges such as non-stationarities in the amplitude and time to peak of hemodynamic responses to a stimulus.

Publicly available datasets were analyzed in this study. This data and the code for HRF ground-truth generation can be found here: https://www.nitrc.org/frs/?group_id=1071.

The studies involving human participants were reviewed and approved by Boston University IRB. The participants provided their written informed consent to participate in this study.

XL and NG collected the first and second datasets, respectively. AL and MY analyzed the data, prepared the figures, and wrote the report. DB reviewed the report. All authors confirmed the final version of the report.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. The authors declare that this study received funding from Facebook. The funder was involved in the decision to submit the paper for publication.