Wearable functional near‐infrared spectroscopy for measuring dissociable activation dynamics of prefrontal cortex subregions during working memory

Abstract

The prefrontal cortex (PFC) has been extensively studied in relation to various cognitive abilities, including executive function, attention, and memory. Nevertheless, there is a gap in our scientific knowledge regarding the functionally dissociable neural dynamics across the PFC during a cognitive task and their individual differences in performance. Here, we explored this possibility using a delayed match‐to‐sample (DMTS) working memory (WM) task using NIRSIT, a high‐density, wireless, wearable functional near‐infrared spectroscopy (fNIRS) system. First, upon presentation of the sample stimulus, we observed an immediate signal increase in the ventral (orbitofrontal) region of the anterior PFC, followed by activity in the dorsolateral PFC. After the DMTS test stimulus appeared, the orbitofrontal cortex activated once again, while the rest of the PFC showed overall disengagement. Individuals with higher accuracy showed earlier and sustained activation of the PFC across the trial. Furthermore, higher network efficiency and functional connectivity in the PFC were correlated with individual WM performance. Our study sheds new light on the dynamics of PFC subregional activity during a cognitive task and its potential applicability in explaining individual differences in experimental, educational, or clinical populations.

Practitioner Points

• Wearable functional near‐infrared spectroscopy (fNIRS) captured dissociable temporal dynamics across prefrontal subregions during a delayed match‐to‐sample task.

• Anterior regions of the orbitofrontal cortex (OFC) activated first during the delay period, followed by the dorsolateral prefrontal cortex (PFC).

• PFC disengaged overall after the delay, but the OFC reactivated to the test stimulus.

• Earlier and sustained activation of PFC was associated with better accuracy.

• Functional connectivity and network efficiency also varied with task performance.

Wearable high‐density functional near‐infrared spectroscopy captured neural activation dynamics across prefrontal subregions during a delayed match‐to‐sample task. Stimulus onset activated the orbitofrontal cortex (OFC) first, followed by the dorsolateral prefrontal cortex (dlPFC); after the delay, the entire PFC deactivated. These neurocognitive markers and PFC network properties differed between match and nonmatch trials and were correlated with individual performance.

1. INTRODUCTION

Measuring brain activity with high spatial and temporal resolution is crucial to our understanding of the neural mechanisms underlying cognitive performance and impairment. For the past several decades, functional magnetic resonance imaging (fMRI) and electroencephalography (EEG) have maintained their status as the most widely used techniques for measuring task‐based activity in the human brain. However, despite the high spatial localizability of fMRI and the high temporal resolution of EEG, there are limitations to both, such as sensitivity to movement, lack of portability, and inaccessibility for certain populations, making them difficult to use in a wide range of settings.

Functional near‐infrared spectroscopy (fNIRS) indirectly measures brain activity by quantifying changes in the concentration of (de)oxyhemoglobin near the cortical surface through optical imaging (Ferrari & Quaresima, 2012; Steinbrink et al., 2006), producing a signal that is comparable to the blood oxygenation level dependent (BOLD) response as a proxy for neural activation but with higher temporal resolution than fMRI (Liu et al., 2016; Pinti et al., 2020). In addition, the fNIRS signal, which is derived from the attenuation of light from the source to the detector, provides an advantage in spatial resolution compared to EEG, particularly in systems with high‐density channels. In cognitive research, fNIRS has been particularly useful in studies of patient populations and infants, due to its wearability and its robustness to motion artifacts (Baddeley, 2012; Obrig, 2014; Pinti et al., 2020; Quaresima et al., 2012; Wilcox & Biondi, 2015). The recent development of wireless systems with a high‐density array of channels and motion sensors (Frijia et al., 2021; Shin et al., 2017; Vidal‐Rosas et al., 2023; Yaqub et al., 2020) has made it possible to precisely localize the fNIRS signal to specific subregions of the brain, while maintaining a high temporal resolution that may be advantageous for tests of cognition. Leveraging such techniques could open new doors for neurocognitive research that can be applied in various naturalistic settings.

In this study, we used a wearable fNIRS system to explore temporal dynamics and individual variance in neural activation measured during a delayed match‐to‐sample (DMTS) working memory (WM) task. WM is a basic cognitive function that enables us to keep information “in mind” and plays a critical role in other higher‐order cognitive processes such as decision‐making, reasoning, and episodic memory (Baddeley, 2012; Chai et al., 2018). Individual WM performance is a behavioral marker for a range of neurological disorders and is applied as a standard measure of cognitive ability in many experimental, educational, or clinical populations.

Decades of research investigating the neural substrates of WM have implicated the prefrontal cortex (PFC) as playing several key roles, including the active maintenance of internal representations, manipulation of information, and executive control (Baddeley, 2003; Chai et al., 2018; Constantinidis & Klingberg, 2016; D'Esposito, 2007; Lara & Wallis, 2015). Meta‐analysis studies of fMRI data suggest a possible functional dissociation across lateral PFC subregions depending on the particular WM task, memory load, or the nature of the stimulus (Emch et al., 2019; Nee et al., 2013; Wagner & Smith, 2003). For instance, while both the dorsolateral PFC (dlPFC) and ventrolateral PFC (vlPFC) have been implicated in memory maintenance, only the dlPFC is said to be engaged in the active manipulation of information in WM (D'Esposito et al., 1999; Narayanan et al., 2005; Petrides et al., 2002; Rowe et al., 2000; Rypma & D'Esposito, 1999; Wagner et al., 2001). The orbitofrontal cortex (OFC), although most well‐known for its cognitive role in value‐based decision‐making and predictive inference (Riceberg & Shapiro, 2017; Wallis, 2012; Wang et al., 2020), is also highly involved in various aspects of WM such as stimulus evaluation, maintenance, and manipulation (Barbey et al., 2011; Cunningham et al., 2009, 2011; Petrides et al., 2002). Finally, the frontopolar cortex (FP), the most anterior part of the human brain, is associated with task‐related subgoal processing, confidence judgments, and integration of information (Braver & Bongiolatti, 2002; Burgess et al., 2007; Green et al., 2006; Kim et al., 2015; Nyberg et al., 2003; Yokoyama et al., 2010).

Despite evidence of a functional dissociation of PFC subregions with respect to WM (D'Esposito et al., 1999; Narayanan et al., 2005; Petrides et al., 2002; Rowe et al., 2000; Wagner et al., 2001), the low temporal resolution of fMRI has presented a challenge to observing the dynamics of these specialized subregions over the course of a single trial, especially across different task conditions (e.g., across the match and non‐match trials of a DMTS task). On the other hand, while EEG studies have provided some insight as to the temporal dynamics of WM through the comparison of event‐related potentials or oscillatory synchrony across the parietal and frontal lobes (Dong et al., 2015; Onton et al., 2005; Sauseng et al., 2010), it is both difficult to spatially localize EEG signals to subregions of the PFC and to readily connect the neural oscillatory markers to past fMRI studies on neural activation.

Here, we aimed to fully leverage the advantages of fNIRS to better characterize the recruitment of dynamic neurocognitive processes during the execution of a DMTS task and investigate their contribution to individual differences in performance and cognitive ability. Thus far, several fNIRS studies on WM have reported patterns of neural activity similar to that of fMRI studies and have revealed meaningful population‐specific differences (e.g., between patient or age groups) (Ehlis et al., 2008; Fishburn et al., 2014; Herff et al., 2014; Perlman et al., 2014; Shin et al., 2017). Nevertheless, dynamic changes in activation across the PFC subregions during the course of a cognitive task have yet to be investigated in detail.

We used a wearable frontal lobe fNIRS system with a high‐density channel array ‐ NIRSIT (OBELAB, Korea) ‐ which covers the lateral PFC with 48 main channels with a sampling rate of about 8 Hz and has been used actively in clinical research (Jang et al., 2021; Lee et al., 2021; Park et al., 2023; Yang et al., 2019). In addition, applying a searchlight‐based probabilistic mapping of fNIRS channels on the standard Montreal Neurological Institute (MNI) space (see Section 2) allowed us to get enough spatial resolution to dissociate neural signals originating from distinct subregions of the PFC.

The temporal dynamics across frontal lobe subregions were characterized and interpreted in terms of the various cognitive processes recruited across a trial of the DMTS task (i.e., during the memory delay and retrieval periods). Furthermore, to gain insight into the functional role of these subregions in explaining individual cognitive ability, we compared differences between high‐performing and low‐performing participants. In addition to the activation dynamics, we characterized the network properties of the PFC (Duan et al., 2012; Lu et al., 2010) that correlated with individual ability.

2. METHODS

2.1. Neural activity measurement using fNIRS

2.1.1. fNIRS device settings and data acquisition

Neural activity was measured using NIRSIT (OBELAB, Korea), a wearable fNIRS device with 24 light sources and 32 detectors covering the entirety of the forehead (Figure 1a). The main 48 channels were constructed by pairing sources and detectors at a 3‐cm horizontal distance (Figure S1A,B). An additional 68 channels were used for supplementary analyses (20 channels with 3‐cm source–detector distance along the vertical axis, and 48 channels with 3.35‐cm source–detector distance).

FIGURE 1.

Prefrontal subregions covered by the functional near‐infrared spectroscopy (fNIRS) system and neural data analysis pipeline. (a) Localization of the 48 main channels in standard MNI space (red points). Color‐coded Brodmann areas (BAs) across the lateral prefrontal cortex (PFC) estimated for the fNIRS channels (see Section 2.1.3). (b) Frontal subregional localization of the main channels. (c) Overall preprocessing, quality control, and neural data analysis pipeline used in this study. MBLL, modified Beer–Lambert Law; SNR, signal‐to‐noise ratio.

The NIRSIT device was placed on the participants' head according to the conventional 10–20 system (Singh et al., 2005) by using the reference points marked on the device. To minimize discomfort during the experiment, the device was removed from the head for a few minutes at the end of each run, and a quality check was performed prior to starting each run of the task in order to ensure that the optodes were well‐situated with a signal‐to‐ratio (SNR) higher than 30 on at least 80% of the main channels.

2.1.2. Preprocessing and quality control

The entire preprocessing, quality control, and analysis pipeline are summarized in Figure 1c. First, each channel's raw optical intensity data from each wavelength (780 and 850 nm, under 5 mW of power) were filtered using a fourth‐order Butterworth filter with a 0.01–0.5 Hz frequency range to exclude instrumental and physiological noise (Pinti et al., 2019). The resulting intensity measures were converted to optical density (OD) values according to the following formula:

$${\mathrm{OD}}_{\mathrm{ch}}=-\log \left(I/I_0\right),$$

where I is the raw intensity value and I ₀ is the average intensity value from each channel for each run of the task. Finally, the OD values were converted to a measure of the concentration of oxyhemoglobin (and deoxyhemoglobin) according to the modified Beer–Lambert Law (Huppert et al., 2009) and normalized for each channel for each run.

We also conducted various control analyses aimed at eliminating non‐brain signals from our dynamics and functional connectivity (FC) analyses by (1) regressing out translational (x, y, and z) and rotational (pitch, roll, and yaw) movements measured from the NIRSIT device's internal gyrometer and accelerometer and by (2) regressing out signals from the shortest‐distance channels provided by our device (source–detector pairs located diagonally at 2.12 cm, Figure S1A). For short‐channel signals, it is usually recommended that channels with a source‐detector distance less than 1.5 cm are used (Wyser et al., 2020); however, because this was not feasible with the current apparatus, we performed the regression using our 2.12 cm channels and provided the results in the Supporting Information section of the paper.

We also implemented a quality control process by sequentially conducting channel‐level, run‐level, and subject‐level rejections using the following criteria: If the average raw intensity value of the channel was too low (below 10) or the SNR from the 5‐s calibration period was under 30, the channel was rejected for that run. The SNR was calculated as follows:

$$\mathrm{SN}{\mathrm{R}}_{\mathrm{ch}}=-10\log {\left(\sigma /\mu \right)}^2,$$

where $\mu$ is the average signal during calibration and $\sigma$ is the standard deviation. If more than 40% of the channels were rejected due to signal quality, the entire run was rejected; and if two out of the three runs were rejected, the entire subject was excluded from the analysis. According to these criteria, two subjects were excluded from the neural data analysis, and four subjects had one run rejected from their dataset. The raw and processed data are available online (https://data.mendeley.com/datasets/4p7952v4vc).

2.1.3. Searchlight‐based probabilistic mapping of fNIRS channels to Brodmann areas

To localize the main fNIRS channels within the “standard brain” space and associate them with specific Brodmann areas (BA), we applied a searchlight‐based probabilistic mapping method (Figure S1C). A previous study conducted by Sato et al. (2013) performed a simultaneous NIRS‐MRI experiment, to localize the fNIRS channels for the HITACHI NIRS system (Hitachi Medical Corporation, Japan) on the MNI space. For our own study, we first overlapped the NIRSIT main channels on the NIRS array used in Sato et al. and calculated the MNI coordinates through linear interpolation (Table S1). Then, we set a 2 cm diameter searchlight sphere for each channel, and the proportion of BA overlapping with the coordinates of the sphere was calculated based on the cytoarchitectonic BA map (Damasio & Damasio, 1989). Our regions of interest (ROIs) were as follows: OFC (BA11), FP (BA10), dlPFC (BA9 and 46), and vlPFC (BA45). BA44 was excluded due to the low coverage by our fNIRS channels due to its lateral location. In defining our ROIs, we assigned the top four channels with the highest assigned probability to each subregion (channels 32, 46, 16, 29 for OFC; channels 23, 27, 22, 26 for FP; channels 19, 20, 17, 18 for dlPFC BA9; channels 34, 36, 2, 8 for dlPFC BA46; channels 38, 44, 6, 10 for vlPFC; see Figures 1b and S1B).

2.2. Behavioral task design

In order to measure the temporal dynamics of the neural activity during WM, we used a DMTS task design (Figure 2a). In each trial, participants fixated on a cross (“fixation”, 7 s) while waiting for the sample stimulus (two‐by‐two array with four bars oriented at 0°, 45°, 90°, or 135°) that appeared on screen for 0.5 s. Following a delay of 6 s (“delay period”), they were presented with the test stimulus (duration 0.5 s) and given 3 s to decide whether or not the sample and test stimuli were identical, after which they entered their response by pressing “z” or “m” on the keyboard (match/nonmatch, respectively). If they responded “nonmatch,” they were then prompted to indicate which bar in the array was different from the sample stimulus. There were three different types of trials: match (test stimulus was identical to the sample), nonmatch easy (one of the bars in the test stimulus was rotated 30° either clockwise or counter‐clockwise), and nonmatch hard (one of the bars was rotated 15°). Each participant performed 90 trials, divided into three runs, with a short break between each run. There were 30 trials of each trial type, presented in a quasi‐random order such that the same type did not appear three times in a row. Behavioral accuracy was measured as the proportion of correctly answered trials (i.e., overall or for each trial type).

FIGURE 2.

Delayed match‐to‐sample (DMTS) task paradigm and behavioral results. (a) Illustration of a single trial of the DMTS task design, consisting of fixation, delay, decision, and response periods. Visual arrays consisting of four lines with different orientations were used as the visual stimuli. After the fixation period, participants were shown one of three types of test stimuli: an array of bars identical to the sample stimulus (match trial) or an array with one of the bars rotated 30° (nonmatch easy trial) or rotated 15° (nonmatch hard trial). In the example shown, the top right bar is rotated. When participants selected “nonmatch” as their response, they were prompted to identify which was the bar that changed orientation (“detail”). The duration of response period was calculated as the mean ± standard deviation across subjects. (b) Behavioral results including overall accuracy and accuracies for each trial type. All accuracy measures were significantly above chance level (all p‐values <.001; match: t ₂₅ = 33.50; nonmatch easy: t ₂₅ = 20.49; nonmatch hard: t ₂₅ = 6.19) but not at ceiling (all p‐values <.001; match: t ₂₅ = −3.94; nonmatch easy: t ₂₅ = −5.46; nonmatch hard: t ₂₅ = −10.57). * and *** denote p < .05 and p < .001, respectively.

Participants were 26 healthy adults (mean ± std: 21.69 ± 2.19 years of age, 12 female), right‐handed, with normal or corrected‐to‐normal vision, with no known neurological disorders. Participants gave written informed consent and received monetary compensation for participation in the study. All experimental procedures were approved by the Institutional Review Board (IRB) at KAIST, Daejeon, Republic of Korea (IRB# KH‐2018‐25).

2.3. Neural activation dynamics analysis

2.3.1. Temporal dynamics analysis

To estimate the temporal dynamics of the neural activity during the DMTS trial, fNIRS signals were averaged across trials for every subject. First, the neural activation was shifted forward 5 s to adjust for the BOLD response which peaks at around 5 s after stimulus onset (Huppert et al., 2006; Machado et al., 2021). Then, the signal was z‐normalized with respect to the fixation period (−4 to 0 s) as a baseline for each trial, where t = 0 was set to the onset of the sample stimulus. The first and last trials for every run were excluded from the analysis, and any outliers (datapoints exceeding three median absolute deviations) were excluded from the analysis. This was less than 10% of the sampled data points. When comparing the dynamics among different trial types, only the data from the correct trials were used.

2.3.2. General linear model

We used a general linear model (GLM) to systemically quantify the neural activation characteristics in each period of the task (Tak & Ye, 2014). We constructed a design matrix for each subject by using a boxcar function‐based regressor convolved with the canonical hemodynamic response function (HRF) across the fixation, delay, and decision periods of the DMTS task. The six movement signals from the NIRSIT gyrometer (x, y, and z) and accelerometer (pitch, roll, and yaw) were also added as nuisance regressors. The beta coefficient for the design matrix was estimated by conducting a GLM (MATLAB 2020a glmfit) for each subject, and second‐level statistics were calculated for each channel or ROI. For each channel, a family‐wise error correction was performed with other channels that shared either a source or detector.

2.3.3. Temporal decoding of fNIRS channel ensembles

To examine the across‐subject consistency of the neural dynamics, we performed a subject‐level decoding analysis on the fNIRS data (Figure S5). We calculated the decoding accuracy for each time window (2‐s windows with a 1‐s gap between them) on each channel ensemble, on each subject, and then obtained the temporal dynamics of the decoding accuracy for each channel ensemble. Higher decoding accuracy at certain timepoints indicates that the ensemble is better able to discriminate it from the time window before or after it. To measure the amount of change in brain activity across time, we used the slope of the activation across the 2‐s time window for every trial as the relevant feature for each channel used for decoding. The slope is a linear term of the fitted linear equation between time and the activation signal, estimated by the least square method. Channel ensembles were constructed by grouping together six adjacent channels (two vertical by three horizontal), in order to get a sufficient number of features for the decoding analysis. A linear Support Vector Machine (SVM) was used for each ensemble and decoding window; to identify the across‐subject consistency for the decoding dynamics, we applied a leave‐one‐subject‐out cross‐validation paradigm, using the trials from one subject as the test set and trials from the other subjects as the training set, repeated for all subjects in the sample. A higher decoding accuracy implied that there is a greater change in neural response at that given time for that ensemble.

2.3.4. Functional connectivity analysis and graph measures

We conducted FC analyses to investigate potential individual differences in coordinated activity between PFC subregions. Given that FC analysis is typically performed in a lower frequency range and longer time windows (Xu et al., 2015; Zhang et al., 2010), we applied a bandpass filter of 0.01–0.1 Hz and utilized the signal from the entire duration of the task, specifically for the FC analysis. As the rapidly changing trial‐level signal difference may not be captured with the lower frequency range, our focus was on individual differences in FC that correlated with WM performance. FC was quantified with z‐transformed (Fisher's transform) Pearson's correlations for each channel pair (or ROI pair) and subsequently rank‐normalized for each individual.

For the network analysis, we constructed functional networks at the channel‐level and compared graph measures across subjects. Functional networks were created as binarized graphs with channels as nodes and binary values from the FC threshold as edges. Multiple graphs were constructed based on different FC thresholds, or “sparsity,” which is defined as the proportion of edges used among the number of all possible edges (Niu et al., 2012, 2013). The range of sparsities was chosen from 0.01 to 0.5 (with step size 0.01) in order to secure the minimum number of edges, taking small‐worldness into consideration (Niu et al., 2012). We calculated representative graph measures—average local efficiency, clustering coefficient, and modularity (Rubinov & Sporns, 2010)—for each sparsity value (also for random graphs as a control analysis), in order to validate the graph measures reflecting small‐worldness. Random graphs were constructed by shuffling the edge values, and average graph measures were obtained by shuffling the channel indices for each subject 1000 times.

2.3.5. Group‐level comparison and behavioral correlation analysis

In order to assess individual differences in the neural activation dynamics, we divided the subjects into subgroups by the median accuracy value of match and nonmatch trials. The number of subjects in each subgroup was 9 versus 8 (high‐ vs. low‐performers) in the match trials and 11 versus 11 in the nonmatch trials. Group differences were analyzed by applying permutation tests (10,000 iterations, shuffling group labels) for the temporal activation dynamics, GLM beta‐values, and FC results.

3. RESULTS

3.1. Behavioral results

Participants (n = 26, mean age 21.69 ± 2.19 years, 12 female) performed the task with high overall accuracy (0.84 ± 0.08). A repeated‐measures ANOVA revealed a main effect of trial type on task accuracy (F _2,46 = 55.25, p < .001, η ² = 0.70); post hoc analyses also showed that match trials had highest accuracy and nonmatch hard trials were the most difficult (Figure 2b; match versus nonmatch easy: t ₂₅ = 2.71, p _corr = .018, Cohen's d = 0.52; match versus nonmatch hard: t ₂₅ = 8.78, p _corr <.001, d = 1.67; nonmatch easy versus hard: t ₂₅ = 7.87, p _corr <.001, d = 1.45). Average accuracies were 0.95 ± 0.07 (match), 0.89 ± 0.10 (nonmatch easy), and 0.68 ± 0.15 (nonmatch hard), and all conditions were above chance level (0.5) but below ceiling (p < .05 for all statistical tests). In the nonmatch trials, subjects successfully reported the position of the changed bar with an accuracy of 0.96 ± 0.06 on easy trials and 0.93 ± 0.08 on hard trials. Accuracies in the nonmatch easy and hard trials were significantly correlated (Kendall's 𝜏 = 0.38, p = .01), while match and nonmatch trials were not (𝜏 = 0.23, p = .14), so these two trial types were separated in further subject‐level analyses. Sex and age did not have a significant effect on overall accuracy (age: F _2,46 = 0.69, p = .51; sex: F _2,46 = 0.05, p = .95).

The average response time (RT) of match trials was 0.73 ± 0.27 s and RT of the “detail” response, in which participants reported which element was different, was 0.54 ± 0.21 s. After rejecting outlier trials and subjects that exceeded three median absolute deviations (one subject rejected), RT was significantly different across trial types, despite being given 3 s in the pre‐response decision period. A repeated‐measures ANOVA showed a main effect of trial type on RT (F _2,44 = 5.35, p = .008) with a significant RT difference between match and nonmatch trials in the post hoc analysis (easy: t ₂₄ = 2.82, p = .01; hard: t ₂₄ = 2.50, p = .02). RTs for both the response and detail periods were not different between nonmatch easy and hard trials (response RT: t ₂₄ = 0.01, p = .993; detail RT: t ₂₄ = 0.21, p = .833).

3.2. Spatiotemporally dissociable PFC neural dynamics

We measured neural activity using a high‐density fNIRS system (NIRSIT; OBELAB, Korea) and performed the analysis according to the pipeline described in Figure 1c. The main results are presented with the main 48 channels (3‐cm source–detector distance) and oxyhemoglobin data, but converging results derived from 68 additional channels and deoxyhemoglobin data can also be found in the Supporting Information. To localize the main channels of the fNIRS on the standard MNI space, a searchlight‐based probabilistic mapping procedure (Figure S1C) was applied, and each channel was assigned to one of the following subregions of the PFC: OFC (BA11), FP (BA10), dlPFC (BA9 and 46), and vlPFC (BA45) (see Figures 1b and S1B).

3.2.1. Temporal dynamics of neural activation varied across PFC subregions during the DMTS task

The neural activation measured by fNIRS changed synchronously with the progression across the phases of the DMTS task and showed dissociable patterns across PFC subregions. As shown in Video S1 and Figure 3a, activation of the PFC increased in the delay period following the presentation of the sample image, first in the ventral channels (including those localized to the OFC; Figure 3b; 0.5–6.5 s, t ₂₃ = 3.16, p = .004, d = 0.62) and then in the dorsal channels (including those localized to the dlPFC BA9; Figure 3b; 2.5–3.0 s, t ₂₃ = 2.18, p = .04, d = 0.43). When directly comparing the activation (uniformly normalized between sample and test stimuli presented), the OFC activation was also significantly higher than the dlPFC (t ₂₃ = 3.22, p = .004) at the moment the sample stimulus was presented. The slope of the activation dynamics also showed a positive increase in activity in the OFC first, followed by an increase in the dlPFC (2.6 s of difference, Figure 3c). When considered alongside the time course of the DMTS task, this activation pattern may indicate an initial representation and evaluation of the sample stimulus by the ventral regions of the PFC followed by the maintenance or manipulation of the memory representation by the dlPFC during the delay period.

FIGURE 3.

Temporal dynamics of neural activation across the prefrontal cortex (PFC) subregions over a trial of the delayed match‐to‐sample (DMTS) task. (a) Snapshots from Video S1 of the neural activation at 1 s (early delay), 2.5 s (mid delay), 8 s (decision), and 11.5 s (response), from the onset of the trial epochs. (b) Temporal dynamics of neural activation across the DMTS task for PFC subregions, baseline‐normalized with respect to the pre‐stimulus fixation period. Phases of the behavioral task are shown in parallel using gray (fixation), red (delay), and blue (decision) shading. The shading around the lines denotes the standard error of the mean and the red bars above the graph indicate statistical significance against baseline (p < .05). (c) Temporal dynamics of the derivative of neural activation, measured by the slope of activation change estimated from 2 s windows. dlPFC, dorsolateral PFC; OFC, orbitofrontal cortex; vlPFC, ventrolateral PFC.

After the initial activation phase, most of the channels were deactivated (Figure 3a,b; during 7–10 s; FP: t ₂₃ = −4.24, p < .001, d = 0.84; dlPFC BA9: t ₂₃ = −3.12, p = .005, d = 0.62; dlPFC BA46: t ₂₃ = −2.56, p = .017, d = 0.51; vlPFC: t ₂₃ = −3.05, p = .006, d = 0.60). These results lend themselves to the possibility that the PFC was activated throughout the early part of the trial and then disengaged in the latter part of the task, perhaps when the cognitive load was resolved by the appearance of the test stimulus. Interestingly, unlike the other PFC subregions, the OFC failed to show the post‐delay deactivation (time 7–10 s, t ₂₃ = 0.22, p = .83) but instead became reactivated once again (Figure 3a,b; time 11.5–15 s, t ₂₃ = 2.91, p = 0.008, d = 0.58). Because this second peak showed a clear difference across trial types (see the Section 3.2.3), the OFC may serve a distinct functional role depending on the detection of a match/non‐match test stimulus.

3.2.2. Confirmatory analyses of neural activation dynamics using GLM and control analyses

Next, we performed several confirmatory analyses to identify subregional specificity in neural activation during the memory delay and decision periods. First, a GLM was conducted to investigate neural activity while factoring out movement‐related changes in the signal. When holding the memory in mind during the delay period, the OFC and dlPFC showed significant activation (Figure 4a, delay period; OFC: t ₂₃ = 4.43, p < .001, d = 0.87; dlPFC BA9: t ₂₃ = 2.63, p = .02, d = 0.52; dlPFC BA46: t ₂₃ = 3.07, p = .005, d = 0.61); then, after the test stimuli were presented, FP and dlPFC BA9 were significantly deactivated (Figure 4b, decision period; FP: t ₂₃ = −3.87, p = .001, d = 0.76; dlPFC BA9: t ₂₃ = −3.79, p = .001, d = 0.75). A similar activation pattern was also observed in the remaining channels as well (Figure S4A): first, the lateral channels were activated, while the medial channels including FP regions were not; afterwards, the medial channels were significantly deactivated. To further confirm the robustness of the temporal dynamics results, we extended the trial epoch to include the next trial: successive trials showed identical neural dynamics pattern (Figure S3B), even when the movement signals (Figure S3C) and 2.12 cm‐channel signals (Figure S3D) were factored out.

FIGURE 4.

Neural activation across the prefrontal cortex (PFC) as estimated by the general linear model during the delayed match‐to‐sample task. Second‐level analysis from beta coefficients estimated for each subject in the (a) delay and (b) decision periods. The yellow circles on the NIRS channel plot (top row) indicates p < .05 (those that survived both false discovery rate (FDR) correction by all channels and Bonferroni correction by source–detector assignment, see Section 2.3.2.). In the bar graph grouping together channels into PFC subregions (bottom row), error bars indicate standard error of the mean, and asterisks denote FDR‐corrected p's < .05 based on regions of interest (*p < .05, **p < .01, and ***p < .001). dlPFC, dorsolateral PFC; FP, frontopolar cortex; OFC, orbitofrontal cortex; vlPFC, ventrolateral PFC.

In order to confirm across‐subject consistency of the temporal dynamics results without applying the 5‐s adjustment for the hemodynamic response, we tested the task phase decodability from ensembles of fNIRs channels (Figure S5A,B). For each channel ensemble, we set the decoding window size to 2‐s (with a 1‐s gap between windows) and calculated the slope of the activation for each trial (as a measure of the change in activation across time across each time window). The subject‐level decoding accuracy was measured using an SVM with leave‐one‐subject‐out cross‐validation for each ensemble.

As the delay period started, the ventral ensembles showed significant decodability first, which then propagated to the other ensembles including the dorsal regions and, finally, the whole PFC (Figure S5C, first row). This result is consistent with the activation dynamics results in the previous section, with the activation starting first in the OFC and propagating to the dlPFC across the delay period of the DMTS task. As time passed, both the dorsal/superior and ventral/inferior ensembles showed significant decodability, followed by all other channels in the latter part of the trial (Figure S5C, second row). These patterns of decoding, using activation slope changes across a sliding time window, converge directly upon the activation dynamics analyses, even without applying a hemodynamic response delay, and confirm that there is significant consistency in activation dynamics across subjects.

3.2.3. Prefrontal subregions respond differently in match and nonmatch trials

Different trial types showed different activation patterns depending on the kind of test stimuli given, reflecting dissociable cognitive roles of PFC subregions at the decision period. First, while the match trials and nonmatch easy trials showed a second peak of OFC activation, the nonmatch hard trials did not, even when only the correct trials were analyzed (Figure S2B,C; time 11–13 s, match: t ₂₃ = 2.62, p = .015; nonmatch easy: t ₂₃ = 2.95, p = .007; nonmatch hard: t ₂₃ = 0.36, p = .72). Given the evaluative function of the OFC (Padoa‐Schioppa & Cai, 2011), these results potentially reflect the higher informational value of the test stimulus in the match and nonmatch easy trials, compared to the nonmatch hard trials.

In contrast to the OFC, the dlPFC showed higher activation in the match trials than the nonmatch trials (Figure S2B; time 9–11.5 s, t ₂₃ = 2.97, p = .007) with no difference between easy and hard trials (Figure S2C; t ₂₃ = 0.95, p = .35). The cognitive processes underlying match and nonmatch trials may differ in that, while nonmatch trials simply require a change detection in one part of the stimulus, match trials require recognition of the entire array. The sustained dlPFC activation in the match trials may indicate a memory comparison process, potentially involving the utilization of a mnemonic strategy (Mansouri et al., 2007; Müller et al., 2002; Wagner et al., 2001). The FP showed only slight differences in activation dynamics during the latter part of the trial, possibly reflecting memory confidence: activation was higher on match than nonmatch trials (Figure S2B; time 11.8–13.5 s; t ₂₃ = 2.30, p = .031) and higher on easy trials than hard trials (Figure S2C; time 11.7–12.5 s; t ₂₃ = 2.19, p = .039).

3.3. Individual differences in performance reflected in frontal lobe dynamics

3.3.1. Better performance is associated with faster and more sustained PFC activity

In order to examine performance‐related individual differences in the activation dynamics, we divided subjects into two groups based on the median accuracies of match or nonmatch trial types and tested for activation differences using permutation tests (shuffling group labels). When comparing between subjects divided by their match trial accuracy, we found a prominent peak timing difference in the delay period (Figure S6A,B; Video S2). High‐performing subjects who made no errors in the match trials showed a faster activation in the early part of the DMTS trial (Figure 5a, comparison across 13 significant channels with average peak timing 3.01 ± 0.77 s versus 5.52 ± 0.43 s, t ₁₅ = −5.11, p < .001, d = 2.51; Figure S7A, comparison across 67 significant channels with average peak timing 2.91 s versus 5.58 s, t ₁₅ = −5.40, p < .001, d = 2.67); however, such a difference in neural activation was not observed between the high‐ versus low‐performing groups in the nonmatch trial (Figure S7A; t ₂₀ = −0.51, p = .62). The combined neural activation measures from the significant channels from Figure 5a clearly indicate that the lateral PFC in the high‐performing participants showed an earlier activation peak, while the peak of the low‐performing participants' activation occurred slightly later (Figure 5b). Looking at all of the participants together without dividing them into groups, we found that match trial accuracy and the peak activation timing were significantly correlated (Figure 5c and S7B; 𝜏 = −0.48, p = .003).

FIGURE 5.

Earlier activation peak and sustained prefrontal cortex (PFC) activity reflect better individual performance. (a) Peak timing difference between high‐ and low‐performers in match trials, with earlier activation in good subjects. The yellow circles indicate statistical significance with p _permute <.05. (b) Activation dynamics of high‐ and low‐performers in match trials, extracted from the significant channels in (A). The colored line above the graph indicates statistically significant difference (p _permute <.05) between the two groups. Phases of the behavioral task are shown in shades of gray: light (fixation), medium (delay), and dark (decision). See Video S2 for channel‐wise dynamics. (c) Correlations between average peak timing and match trial accuracy. (d) Averaged difference of the level of activation across the continuous period of 7–10 s (corresponding to the decision period of the delayed match‐to‐sample working memory task) between high‐ versus low‐performers in the nonmatch trials. High‐performers showed more sustained activation than low‐performers. (e) Activation dynamics of high‐ and low‐performers in match trials, extracted from the significant channels in (d). See Video S3 for channel‐wise dynamics. (f) Correlation between the peak activation level and nonmatch trial accuracy.

When comparing activation dynamics between subjects divided into good‐ and bad‐performer groups according to nonmatch trial accuracy, we found that they differed significantly in the level of deactivation in the latter half of the trial corresponding to the decision‐making phase of the task (Figure S6C,D; Video S3). Good performers in the nonmatch trials showed less deactivation in the decision period across the PFC, particularly in the FP region (Figures 5d,e and S7C; t ₂₀ = 3.08, p = .006 from the significant channels, d = 0.43). FP channels showed significant correlation between participants' nonmatch trial accuracy and their peak activation (Figures 5f and S7D; 𝜏 = 0.34, p = 0.028), as well as the activation coefficient from the GLM analysis (Figure S7F; 𝜏 = 0.36, p = .017).

3.3.2. Functional connectivity and network measures reflect individual ability

We explored whether FC within the frontal lobe during the task can reveal novel dynamic neural markers of individual performance. We measured FC by calculating correlations for each pair of channels (or ROIs), constructing an FC map for every participant, and then correlating it with DMTS match and nonmatch trial performance (Figure 6a; see Figure S8 for control analyses regressing out movement and 2.12 cm channels).

FIGURE 6.

Functional connectivity (FC) and graph measures reflect individual differences in delayed match‐to‐sample working memory performance. (a) Schematic diagram of FC and graph measure analyses with group‐level comparisons and correlations. (b, d) Behavioral relevance of FC shown both in matrix form, with each row/column representing the left/right prefrontal cortex (PFC) subregions (* denotes significant differences between high‐ and low‐performing groups and ° denotes FC between regions), and as a topological network across the PFC with bold lines indicating significant group differences. (c) Correlation between graph measures and match trial accuracy; the plot is drawn for peak correlations among multiple sparsity values (Figure S9B, sparsity 0.14). (e) Correlations between graph measures and nonmatch trial accuracy shown for peak correlations among multiple sparsity values (Figure S9C, sparsity 0.10). BA, Brodmann areas; dlPFC, dorsolateral PFC; FP, frontopolar cortex; OFC, orbitofrontal cortex; ROI, regions of interest; vlPFC, ventrolateral PFC.

First, good performers in the match trials showed higher FC across the PFC subregions, especially between the dlPFC and other subregions (Figures 6b and S8A,C). On the other hand, the correlation between nonmatch trial accuracy and FC was only significant in the interhemispheric connectivity in the FP cortex (lower FC in good performers; Figures 6d and S8B,D; 𝜏 = −0.35, p = .02), which is consistent with the activation dynamics results showing that better performers showed sustained activation of FP regions after the memory delay.

Finally, we constructed a functional brain network with multiple sparsity values (proportion of edges used for the graph construction, out of all possible edges) and then calculated the correlation between graph measures and behavioral accuracies (Figure S9). We found that modularity was positively correlated with match trial accuracy (Figures 6c and S9D; peak sparsity = 0.14, 𝜏 = 0.33, p _corr = .038, p _permute = .020), indicating that the functional PFC network of high‐performing participants showed more segregated communities. Meanwhile, the local efficiency of the graph was correlated with nonmatch trial accuracy (Figures 6e and S9E; peak sparsity = 0.10, 𝜏 = 0.29, p _corr = .053, p _permute = .023), which indicates higher inter‐connectivity between edges of the PFC network in high‐performers.

4. DISCUSSION

4.1. Prefrontal activation dynamics during cognitive processing

The present study extends our current understanding of dissociable functional roles of PFC subregions and demonstrates their rapidly changing activation dynamics during a DMTS task of WM. First, we observed that the anterior part of the OFC is the first to activate, immediately following stimulus onset, and that the dlPFC activated slightly after that. The OFC is well‐known for its role in evaluating external stimuli (Cunningham et al., 2009, 2011; Rolls, 2004), but it is also reported to play an important role in multiple memory processes and in integrating information from other brain regions (Frey & Petrides, 2002; Lara et al., 2009; Nogueira et al., 2017; Stalnaker et al., 2015; Wallis, 2007). The dlPFC, on the other hand, is important for manipulating information in WM (Meyer et al., 2011; Petrides et al., 2002; Rowe et al., 2000; Wagner et al., 2001) and for representing cognitive load (Rypma & D'Esposito, 1999, 2003). Our results add to this literature by showing that the OFC was not only recruited for early processing and assessment of perceptual information (sample stimuli) but also a second time during the sample‐to‐test matching phase. This may reflect the re‐evaluation or monitoring process of the anterior OFC during the decision period (Meyer et al., 2011; Petrides et al., 2002; Rowe et al., 2000; Rypma & D'Esposito, 1999, 2003; Wagner et al., 2001).

After the delay period, most of the PFC subregions were deactivated. Deactivation in neuroimaging studies can be attributed to various causes, including the activation of the default mode network, vascular steal from adjacent brain regions, or relative decrease due to activation in the baseline period (Hayes & Huxtable, 2012; Moraschi et al., 2012). In our study, the observed frontal deactivation quickly recovered back to the baseline level within several seconds and stabilized for the early part of the next trial. This pattern of activation dynamics supports the interpretation that the PFC was moderately engaged starting from the fixation period, retained its activation during the memory delay, and then disengaged in the post‐stimulus period, once the test stimulus was presented.

4.2. Performance‐dependent individual differences in activation dynamics and functional connectivity

Remarkably, the prefrontal dynamics during the DMTS task showed clear performance‐related individual differences. In participants with good match trial performance, the PFC activation was initiated quickly, suggesting that faster assessment of the stimulus and assignment of prefrontal neural resources may be important for high cognitive performance. The fact that this was true only in the match trials may indicate that a more rapid representation of the stimulus may be crucial for its subsequent recognition following a delay. Furthermore, this performance‐dependent early activation might be related to findings from EEG studies that showed that stronger prefrontal event‐related potentials and more synchronized interregional theta oscillations are correlated with WM performance (Dong et al., 2015; Sauseng et al., 2010).

High performance in the nonmatch trials was correlated with sustained activation (less disengagement after the memory period) of the FP cortex, along with high FC and local network efficiency. The FP region has been implicated in WM through its role in subgoal processing, executive control, confident judgment, and the monitoring and integration of memory (Braver & Bongiolatti, 2002; Green et al., 2006; Kim et al., 2015; Yokoyama et al., 2010). Studies have also suggested that FP is critical for the interaction between WM and perceptual attention, for instance in tasks that require that an item is held in WM while attending to another visual goal (Koechlin et al., 1999), or tasks that involve switching between internal and external representations (Burgess et al., 2007). Although several of these functions may be simultaneously at play in our study, the finding of individual differences in FP dynamics in the nonmatch trials may be attributed to sustained cognitive processing even after the delay period (Dove et al., 2000) and possibly reflect the detection of a discrepancy between the sample and test stimuli during the decision period (Pollmann et al., 2006; Pollmann & Manginelli, 2009).

Our findings extend the existing body of research on the neural correlates of individual cognitive ability, based on differences in not only activation across the PFC subregions but also in the speed of neural processing, maintenance of activation, and network properties across the PFC (Dong et al., 2015; Nozawa & Miyake, 2020; Osaka et al., 2003; Rypma & D'Esposito, 1999; Sauseng et al., 2010; Zhang et al., 2013). Moreover, by identifying these neurocognitive markers during a fast‐paced cognitive task using a wearable fNIRS system, we hope that our findings may be utilized in evaluating and monitoring WM function across various patient populations (Baier et al., 2010; Faraza et al., 2021; Jarrold & Towse, 2006; Townsend et al., 2010).

4.3. Limitations

Despite the implementation of a high‐density fNIRS system, there are limitations to how precisely one can localize the source of the neural response, especially in the absence of individual neuroanatomical data. Like most NIRS systems, NIRSIT follows a standard 10–20 system (Singh et al., 2005), and our probabilistic mapping of BA was based on a previous study that localized fNIRS channels using MRI scans (Sato et al., 2013; see Section 2). However, because the Sato et al. study used a different fNIRS system, there is still a lack of absolute certainty about our subregion labels. Especially for the OFC, the fNIRS channels only cover the most anterior part of the OFC and therefore might not represent the activity of the entire OFC. Nevertheless, the temporal dynamics analyses showed clearly dissociable activation patterns across broadly localizable regions of the PFC, at least to the extent that we can confidently distinguish between ventral‐dorsal (e.g., OFC vs. dlPFC) and medial‐lateral (e.g., FP vs. others) subregions.

Second, measuring changes in the oxyhemoglobin concentration as an indirect measure of brain activity suffers from heavy delays compared to the actual neural response (Pinti et al., 2019). We were able to partially address this issue by applying the HRF in our GLM, shifting the signal in time in our activation dynamics analysis, and finding converging evidence in our decoding analysis using the slope of activation change across a time window. Nevertheless, as the subregional difference was derived from their relative timing differences, we hold that the dissociation of their temporal dynamics is still valid.

Finally, the current task design may be too rapid to capture the dynamics of the BOLD signal, leaving room for some uncertainty about whether the dynamics were specifically time‐locked to the stimulus or test stimulus onset. We address this concern by confirming the robustness of the current results using various control analyses and extended epoch analysis, and also validated that partial sets of trials (trial‐type difference or odd/even trial separation) exhibited consistent patterns.

4.4. Conclusion

In this study, we characterized the dissociable temporal dynamics of neural activation across PFC subregions during cognition using a high‐density wearable fNIRS system (NIRSIT). Upon onset of the sample stimulus in a DMTS task, the ventral region (OFC) immediately activated, followed by the dlPFC during the memory delay period. The OFC activated again, coinciding with when the test stimulus appeared, while the PFC disengaged overall during the retrieval period. These results converge with the current understanding on the dissociable functional roles of the PFC subregions and provide novel findings regarding their temporal activation dynamics in a DMTS task, with insights that may be more broadly applicable to other WM tasks. Along with the identification of individual performance‐related markers in temporal dynamics and PFC connectivity, these results have the potential to be utilized in conjunction with wearable optical imaging tools across a wide range of settings, including educational neuroscience, digital healthcare, and applied neuroscience.

CONFLICT OF INTEREST STATEMENT

The authors declare there is no conflict of interest.

Supporting information

Data S1. Supporting information.

Video S1. Supporting information.

Video S2. Supporting information.

Video S3. Supporting information.

Shin, J. H. , Kang, M. J. , & Lee, S. A. (2024). Wearable functional near‐infrared spectroscopy for measuring dissociable activation dynamics of prefrontal cortex subregions during working memory. Human Brain Mapping, 45(2), e26619. 10.1002/hbm.26619

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are openly available in Mendeley Data at https://data.mendeley.com/datasets/4p7952v4vc/1.