Hemispheric Network Dynamics During Auditory Language Comprehension and Its Clinical Implications Regarding Resting-State Functional Magnetic Resonance Imaging

Abstract

Purpose:

With growing interest in modeling neurobehavior, there is increased interest in understanding patterns of functional connectivity (FC) during language processing. Previous research has suggested that static resting-state functional magnetic resonance imaging (rs-fMRI) and task-based functional magnetic resonance imaging (tb-fMRI) may be interchangeable in determining FC in language-related regions of interest. Authors have argued for the elimination of using tb-fMRI assessments in preoperative clinical workup of language mapping. However, given that language exhibits not only 3D spatial attributes but also temporal components, understanding the temporal dynamics is essential in developing adaptive computational models. Thus, the stability of language neural networks during rs-fMRI and tb-fMRI during auditory comprehension was examined in healthy participants.

Method:

Twenty-three participants underwent rs-fMRI and 12 participants underwent tb-fMRI while listening to an auditory description task. Sliding scale time series correlation was used to generate estimates of dynamic FC. We used Spearman correlation with Bonferroni correction to compute statistically significant whole-brain dynamic networks. A total of 59,831 data points (42,159 in rs-fMRI, 10,152 in tb-fMRI, and 7,520 in overlap) were analyzed.

Results:

There was more notable fluctuation in linear correlation over time for dynamic FC networks activated during the task relative to baseline than during rs-fMRI. Differences in dynamic FC were noted in only specific common networks in the left hemisphere during tb-fMRI.

Conclusions:

The temporal dynamics of identical neural networks during rs-fMRI and tb-fMRI are markedly different, especially within the left hemisphere. This may be an important computational model feature that may improve model prediction of clinical outcomes following central nervous system injury.

Language processing requires highly complex concurrent processes that involve spatial–temporal dynamics of neural activity (Canolty et al., 2007; Jarret et al., 2022). The dual-stream model of speech and language processing, which includes a bilateral ventral stream and left-lateralized dorsal stream, suggests that speech is processed concurrently by networks using multiple timescales (Hickok, 2022; Hickok & Poeppel, 2007). The acoustic speech signal is projected to several nodes in two anatomic streams, a dorsal stream and a ventral stream, for spectrotemporal analysis, phonological-level processing, and lexical–semantic–grammatical processing. Then, the phonological, semantic, lexical, and grammatical information is integrated into speech motor functions and representations during speech production and comprehension (Hickok, 2022; Hickok & Poeppel, 2007). Given these spatial and temporal dynamics associated with the acoustic signal neuroprocessing, both spatial connectivity networks and the temporal dynamics of the networks are crucially important for understanding language processing. It is essential to investigate the differences in time-varying changes on neural networks during language tasks because disruption of either the spatial and/or temporal components of neural networks may result in language disorders, including aphasia and neurodegenerative disorders, for example. Indeed, 30%–60% of patients poststroke show communication disorders, such as aphasia, dysarthria, apraxia of speech, and cognitive-communication impairments (Tippett, 2018). Approximately one third of stroke survivors acquire aphasia, which negatively impacts one's employment, social interactions, quality of life, and mental health (Flowers et al., 2016; National Aphasia Association, n.d.). Developing neurocomputational models of spatial–temporal dynamic neuronal functional changes in patients with central nervous system (CNS) injury would be a valuable approach, with theoretical and clinical significance, to identify neurobiological predictors (Y.-N. Chang & Lambon Ralph, 2020).

Recent studies with neurocomputational models have begun to examine functional connectivity (FC; Y.-N. Chang & Lambon Ralph, 2020; Zhu et al., 2023). FC analysis is a noninvasive neuroimaging technique that examines temporal and spatial statistical association in blood-oxygen-level–dependent (BOLD) signal changes among anatomically distinct regions (Friston et al., 1994; Rosch & Mostofsky, 2019). Resting-state FC in functional magnetic resonance imaging (rs-fMRI) allows one to examine spontaneous fluctuations in BOLD signals without performing any tasks (Field & Birn, 2021; Klingbeil et al., 2019; Rogers et al., 2007). Contrary to task-based FC in functional magnetic resonance imaging (tb-fMRI) that requires patients to engage and perform language and/or cognitive tasks during imaging examinations, rs-fMRI can be performed by patients regardless of the severity of their language and/or cognitive deficits (Field & Birn, 2021; Klingbeil et al., 2019). Given this advantage, rs-fMRI has been employed in language research as an alternative to tb-fMRI (Beheshtian et al., 2021; Doucet et al., 2017; Holiga et al., 2018; Lee et al., 2013; Lv et al., 2018; Smitha et al., 2017). Additionally, a majority of previous FC language studies employed static FC approaches. Static FC approaches assume FC stays constant during the imaging examination, although FC is known to fluctuate over time (Calhoun et al., 2014; C. Chang & Glover, 2010; Zhang et al., 2018). On the contrary, dynamic FC approaches, which were developed to investigate brain connectivity changes over time, allow for the examination of temporal fluctuations of FC over the imaging examination (Cohen, 2018; Hutchison et al., 2013; Zhang et al., 2018).

There has been increased interest in determining connectivity correlations during language processing in healthy individuals as well as patient populations including patients with CNS injury. Previous studies have identified FC related to language processing, such as auditory comprehension, speech production, reading, and writing in healthy participants in both rs-fMRI and tb-fMRI (Battistella et al., 2020; Liu & Yang, 2023; Tomasi & Volkow, 2012; Yu et al., 2018). Stroke and neurodegenerative diseases disrupt FC (Meier, 2022). Previous studies have also indicated reduced FC in people with aphasia poststroke (Keator et al., 2021; Klingbeil et al., 2019; Meier, 2022; Tang et al., 2016; Zhu et al., 2025). Additionally, treatment-induced language recovery can be reflected in connectivity changes (Kiran et al., 2015). Despite high variability in therapy responses in the population, there is evidence that shows changes in FC (e.g., increased and/or normalized FC) are correlated with language recovery in people with aphasia (Baliki et al., 2018; Duncan & Small, 2016; Kiran et al., 2015; Kristinsson et al., 2022; Sandberg et al., 2015; van Hees et al., 2014; Vitali et al., 2010). Given these lines of evidence, FC analysis is a promising approach to investigate neurobiological predictors of speech and language therapy outcomes in people with aphasia. Machine learning models with FC patterns may become useful for predicting language functions following CNS injury as well as language recovery trajectories.

Several lines of evidence indicate that there are differences in the characteristics between rs-fMRI and tb-fMRI. For instance, previous studies indicated there is more fluctuation in static FC in rs-fMRI than in static tb-fMRI. There is within- and between-subject FC variability across networks in static rs-fMRI in healthy individuals due to various endogenous and exogenous factors (Barch et al., 2013; Goldstone et al., 2016; Mueller et al., 2013; Sjuls & Specht, 2022; Vaisvilaite et al., 2022). Individual effects (i.e., within-subject variability) in FC in rs-fMRI is highest in temporal, frontal, and parietal association cortex areas in healthy participants (Mueller et al., 2013). In contrast, static tb-fMRI is a more straightforward method to obtain more reliable language mapping (Field & Birn, 2021). Changes in FC in tb-fMRI have increased prediction accuracy of cognitive and language task activations, including an auditory comprehension task in healthy individuals (Cole et al., 2021). For these reasons, some authors argue for synchronization and greater reliability of static FC in language-mediated networks during tb-fMRI than rs-fMRI (Field & Birn, 2021).

Alternatively, some authors argue for the relative stability of static connectivity in language-related neural networks in both rs-fMRI and tb-fMRI (Gonzalez-Castillo & Bandettini, 2018). Cole et al. (2014) indicated that most networks in rs-fMRI were identifiable in tb-fMRI with various cognitive tasks and an auditory sentence comprehension task. Differences in FC patterns are subtle between the two modalities in healthy participants. Also, intrinsic static resting-state FC matrices were highly correlated with those of task-based FC during auditory sentence comprehension in the same participants (Cole et al., 2014). Researchers who have applied the static FC technique using independent component analyses to assess language laterality for epilepsy patients reported consistent spatial connectivity for the activated language networks for both static rs-fMRI and static tb-fMRI data with an auditory sentence comprehension task (Rolinski et al., 2020). They argued for the activation of similar spatial network correlations when using either rs-fMRI or tb-fMRI and suggested that data from the two techniques are interchangeable. Although the evaluation of static connectivity correlations may indicate relative stability of the network among language-related brain regions during rs-fMRI and tb-fMRI, it assumes that the temporal network correlations remain constant over time during language processing.

Importantly, FC during tb-fMRI fluctuates over time (Matsui et al., 2019; Tagliazucchi et al., 2012), and FC dynamics reflect task-related neural networks (Betzel et al., 2020; Hutchison et al., 2013; Kucyi & Davis, 2014; Kurtin et al., 2023; Song & Lee, 2023). Previous studies have reported less correlation changes over time in FC networks during a cognitive task, such as a 2-back working memory task, in rs-fMRI in healthy participants (Chen et al., 2015). Another study reported task-induced FC correlation variability reduction during a visual classification task in healthy participants (Elton & Gao, 2015). However, it is known that dynamic FC correlation patterns are task dependent (Shine et al., 2016). FC correlation variability pattern during language processing should be uniquely different from those during working memory and visual classification tasks. Most of the previous FC studies that investigated language processing employed the static connectivity approach, which assumes network constancy. Thus, network dynamics over time during language tasks in rs-fMRI and tb-fMRI have not been well investigated.

Speech contains complex multiscale structure with shorter to longer timescales, and the brain needs to integrate the wide-ranging temporal information for comprehending speech and language (Norman-Haignere et al., 2022; Overath et al., 2015). The aforementioned dual-stream model for speech and language processing suggests that language processing contains temporal dynamics of neural activity to account for the multiscale structure of speech and language. The multiscale processing model and asymmetric sampling in time theory suggest that language processing relies on temporal information (e.g., segmental, syllabic, phonetic information), which is conveyed by acoustic signals at different timescales in bilateral neural networks asymmetrically. Depending on the temporal information, acoustic signals are concurrently processed at multiple different timescales in bilateral networks concurrently (Poeppel, 2003; Poeppel & Assaneo, 2020). For instance, neuroimaging studies have indicated that, after acoustic signals arrive in the bilateral auditory cortex, rapidly changing acoustic signals were processed in the left hemisphere (e.g., superior temporal gyrus), while slowly changing signals were processed in the right hemisphere (Flinker et al., 2019; Oderbolz et al., 2025; Overath et al., 2015). Given the temporal dynamics fundamental to language processing, FC during language processing in tb-fMRI should fluctuate depending on the temporal information conveyed by acoustic signals. Thus, there should be more change over time in language-related networks during tb-fMRI than in the same networks during rs-fMRI.

In the current study, we evaluated FC network dynamics during an auditory comprehension task in healthy participants. We hypothesized that there would be more fluctuation over time in language-related networks during tb-fMRI than in rs-fMRI given the multi-timescale processing and temporal encoding of linguistic information in language-related bilateral neural networks. Results will be crucial contributing factors for using FC to identify neurobiological predictors of language function and language recovery trajectories in patients following CNS injury.

Method

Participants

We conducted a retrospective analysis of healthy adult participants who were recruited as control subjects from the National Institute of Neurological Disorders and Stroke Epilepsy Service as part of the institutional review board (IRB)–approved clinical research protocol for patients with seizures. They originally were recruited to participate in another study as control subjects (Rolinski et al., 2020). All participants provided written informed consent, and the data acquisition protocol was approved by the IRB of the National Institutes of Health (IRB Protocol #14-N-0061).

Originally, 34 participants were recruited, but nine were eliminated either because of excessive motion (n = 3) or because the subject was under 18 years old (n = 6). Thus, data from 25 healthy participants (14 men and 11 women, median age = 27, interquartile range [IQR] = 23–47, range: 21–63 years) underwent analyses for the current study in the following ways: 23 with rs-fMRI and 12 who performed tb-fMRI using language tasks that were matched for task difficulty. There were 10 participants who underwent both rs-fMRI and tb-fMRI (see Table 1). In order to compare identical network correlation dynamics in the same individuals during rs-fMRI and tb-fMRI, we analyzed a subset of 10 participants who underwent both tb-fMRI and rs-fMRI. Because we expected the language networks to be bilaterally located, we conducted a sample size calculation based on a range of expected mean paired differences in laterality index at a two-sided α = .05 and β = .20 that yielded a sample size of eight subjects. This ensured that the analysis of the 10 paired subjects would generate statistically powered results.

Table 1.

Demographic information of each participant.

No.	Age	Sex	Handedness	Education (years)	Race	Resting rs-fMRI (n = 23)	Task tb-fMRI (n = 12)	Overlap (n = 10)
1	23	Male	Right	16	Other	×	×	×
2	23	Male	Right	16	White	×	×	×
3	23	Male	Right	16	White	×	×	×
4	23	Male	Left	> 16	White	×	×	×
5	23	Female	Right	16	Other	×	×	×
6	29	Male	Right	16	Other	×	×	×
7	47	Female	Right	12–15	African American	×	×	×
8	51	Female	Right	16	White	×	×	×
9	53	Female	Left	12–15	African American	×	×	×
10	63	Female	Right	16	African American	×	×	×
11	21	Male	Right	12–15	White	×
12	21	Female	Right	12–15	African American	×
13	22	Male	Right	16	White	×
14	23	Female	Right	16	White	×
15	24	Male	Right	16	White	×
16	25	Female	Right	> 16	Asian	×
17	27	Male	Left	> 16	White	×
18	31	Female	Right	> 16	African American	×
19	32	Male	Right	> 16	White	×
20	39	Male	Right	> 16	African American	×
21	40	Male	Right	> 16	African American	×
22	55	Female	Right	9–12	African American	×
23	55	Male	Right	> 16	White	×
24	21	Male	Right	9–12	African American		×
25	53	Female	Right	< 8	African American		×

Note. rs-fMRI = resting-state functional magnetic resonance imaging; tb-fMRI = task-based functional magnetic resonance imaging.

Participants were monolingual, native speakers of American English. They did not have current and/or past neurological conditions that may cause any language deficits. Most of the participants were right-handed (88%). There was no statistical difference as a function of handedness between participants who underwent only rs-fMRI and only tb-fMRI (Z = −0.63, p = .53). Demographic information of the participants is summarized in Table 2.

Table 2.

Summary of the demographic information of all participants.

Variable		Resting rs-fMRI (n = 23)	Task tb-fMRI (n = 12)	Overlap (n = 10)
Age (years)	Mdn (IQR)	27 (23–47)	26 (23–52)	26 (23–51)
	Range	21–63	21–63	23–63
Sex	Male	13 (57%)	6 (50%)	5 (50%)
	Female	10 (43%)	6 (50%)	5 (50%)
Race/ethnicity	African American	8 (35%)	5 (42%)	3 (30%)
	Caucasian	11 (48%)	4 (33%)	4 (40%)
	Other	4 (17%)	3 (25%)	3 (30%)
Handedness	Left	3 (13%)	2 (17%)	2 (20%)
	Right	20 (87%)	10 (83%)	8 (80%)
Education (years)	< 8	0	1 (8%)	0 (0%)
	9–12	1 (4%)	2 (8%)	0 (0%)
	12–15	4 (17%)	2 (17%)	2 (20%)
	16	10 (43%)	7 (58%)	7 (70%)
	> 16	8 (35%)	1 (8%)	1 (10%)

Note. rs-fMRI = resting-state functional magnetic resonance imaging; tb-fMRI = task-based functional magnetic resonance imaging; IQR = interquartile range.

Data Acquisition

Imaging acquisition was performed on a 3D Echo Planar on a 3T GE MRI at the National Institutes of Health Nuclear Magnetic Resonance Center using a 32-channel head coil. Imaging parameters were as follows: flip angle = 65, repetition time = 2,000 ms, echo time = 30, voxel size = 3 mm × 3 mm × 4 mm, field of view = 216 × 216 mm, slice thickness = 4 mm, number of slices: 38 whole-brain slices from inferior to superior, R to L: 72 slices, and A to P: 72 slices. During rs-fMRI, participants laid down in the scanner while staring at a cross-hair in the display for the duration of the scan. During tb-fMRI, participants were asked to perform an auditory description decision task (ADDT). The ADDT was administered as part of the clinical examination. The ADDT was designed to identify the broad networks involved in language processing for epilepsy presurgical evaluation (Mbwana et al., 2008). During the task, the participants listened to an auditorily presented six- to eight-word sentence and then decided whether the sentence was correct. For example, participants heard “A large gray animal is an elephant” as a correct sentence and “Spaghetti is something you sit on” as a foil sentence. Participants were instructed to press a button when the sentence was correct. The details of the tb-fMRI procedure have been described elsewhere (Gaillard et al., 2007; Rolinski et al., 2020). The paradigm followed a block design of five epoch cycles where each cycle consists of an alternating block of task and control task. This was repeated for a 30-s duration followed by the control task. During the control task, the participant listened to a reverse speech sample (i.e., a sentence played backward).

Data Preprocessing

Given that language processing networks are interconnected to nonlanguage networks (Fedorenko et al., 2024), we decided against using predetermined language-specific parcellations because they may not capture the complete connectivity associated with language processing (Finn et al., 2014). Therefore, we used whole-brain parcellations in the current study because language processing requires cognitive and perceptual processes.

MRI Preprocessing

Anatomical data preprocessing was conducted to achieve accurate surface parcellation. More specifically, surface parcellation was performed with T1-weighted magnetization-prepared rapid gradient echo imaging (MPRAGE) and fluid-attenuated inversion recovery (FLAIR) images using FreeSurfer (v6.0; Fischl, 2012). Surface labeling was done as in Desikan et al. (2006).

fMRI Preprocessing

The fMRI data were preprocessed using surface-based analysis pipelines with the Analysis of Functional NeuroImages (AFNI) toolbox (v19; Cox, 1996). Pre–steady-state volumes prior to reaching equilibrium magnetization were discarded. First, despiking of the signal was done, after which slice timing correction was performed to synchronize timing across slices. The base volume was registered to the corresponding deobliqued anatomical T1-MPRAGE image. As described in the literature, we aligned all volumes to the base volume to account for head movement during the data recording in order to perform motion correction (RaviPrakash et al., 2021). The white matter and ventricle maps from the FreeSurfer output of anatomical data preprocessing were used as tissue regressors to detect non-BOLD signals in the data. Spatial smoothing with a 4-mm kernel was performed within each volume to further reduce the noise and add up coherent signals locally. De-meaned motion parameters were computed for use in regression, and motion was regressed per individual task run. The motion threshold was set to 0.3 mm, and volumes that exceeded this threshold were censored.

rs-fMRI Preprocessing

The pre–steady-state volumes prior to reaching equilibrium magnetization were discarded during acquisition. AFNI was used to preprocess the data. First, despiking of the signal was done, after which slice timing correction was performed to synchronize timing across slices. The base volume was registered to the corresponding de-obliqued anatomical T1-MPRAGE image. We aligned all volumes to the base volume to account for head movement during the data recording in order to perform motion correction. The white matter and ventricle maps from the FreeSurfer output of anatomical data preprocessing were used as tissue regressors to detect non-BOLD signals in the data. Again, a spatial smoothing procedure, with a 4-mm kernel, was performed within each volume to further reduce the noise. We then added up coherent signals locally. The motion threshold was set to 0.3 mm, and volumes that exceeded this threshold were censored. Finally, bandpass filtering was performed for both rs-fMRI and tb-fMRI data to further isolate the noise with a frequency range of 0.01–0.10 Hz.

FC Network Analysis

The task network was generated by first concatenating the time series of all task trials/blocks in a single subject as shown in Figure 1. Pearson's correlation was then computed to generate the static FC. The associated p values were corrected using familywise error correction techniques, specifically, Bonferroni correction. All significant connections (i.e., connections with p values < .05 after correction) were identified as significant task connections. Of all the networks activated, our functional dynamic analysis included only networks common to all participants. Given our small sample size and to achieve maximum sensitivity, we only included networks that were common in all participants for the comparative analysis of the dynamics of network activation over time in the sample. While conservative, this was to ensure network reliability and reproducibility. Thus, “common network correlations” were defined as those that were statistically significant above baseline in all participants. Only those common networks were included in the statistical analyses. There were a total of 41 common networks identified for intra- and interhemispheric networks (39 in rs-fMRI, 18 in tb-fMRI, and 16 in the paired subset). Figure 2 shows common networks in tb-fMRI across all participants. The “interhemispheric network” represents FC that links regions of interest (ROIs) defined in the Desikan atlas between hemispheres. Similar to the task network, the baseline network was generated via concatenation of baseline trials followed by Pearson's correlation. The baseline correlation values also were analyzed. Data analysis was performed using python (v3.7) scripts on PyCharm IDE.

Figure 1.

Task-related time series extraction. This figure illustrates the method to generate task-related functional connectivity networks by concatenating the time series of all task trials/blocks in a single participant.

Figure 2.

Common functional connectivity in task-based functional magnetic resonance imaging (tb-fMRI) across all participants. This figure shows the common functional connectivity networks that were identified in all the participants in tb-fMRI.

We used a sliding-window methodology to calculate the time series correlation over time to generate dynamic FC from both rs-fMRI and tb-fMRI data. Spearman correlation coefficient analyses with Bonferroni correction were then used to compute statistically significant dynamic FC networks in both rs-fMRI and tb-fMRI. We used 12 timepoints with a fixed length stride of one timepoint for both tb-fMRI and rs-fMRI analyses (Cohen, 2018; Hutchison et al., 2013). The window was then slid across the time series by a fixed length stride (i.e., 1 s), and the correlations were computed again. This process was repeated to cover the entire time series. This yielded a total of 47 timepoints for both rs-fMRI and tb-fMRI. Spearman correlation coefficient was then computed across each pair of ROIs for each of the timepoints, resulting in the dynamic FC.

The number of identified common networks in intra- and interhemispheric networks and timepoints in the participants included a total of 42,159 data points (39 networks × 47 timepoints × 23 participants) in rs-fMRI, 10,152 data points (18 networks × 47 timepoints × 12 participants) in tb-fMRI, and 7,520 overlapping data points (16 networks × 47 timepoints × 10 participants) that were analyzed for the current study (see Figure 3).

Figure 3.

Study procedures flowchart. This figure illustrates the flowchart of the study procedure. Data from 23 participants who underwent rs-fMRI, 12 participants who had tb-fMRI, and 10 participants who underwent both rs-fMRI and tb-fMRI were analyzed (a total of 59,831 data points) for the current study. rs-fMRI = resting-state functional magnetic resonance imaging; tb-fMRI = task-based functional magnetic resonance imaging.

Statistical Analysis

Simple descriptive statistics and frequency distributions described participants' demographics. Categorical data were compared between groups using Fisher's exact tests or tests for trend for ordinal data. Continuous data were compared using the nonparametric Wilcoxon rank sum test. We used linear mixed-effect modeling for repeated measures to test FC linear trends over time in common connections in the left hemisphere, right hemisphere, and interhemisphere and in three groups: (a) 23 participants who underwent rs-fMRI, (b) 12 participants who had tb-fMRI, and (c) 10 participants who underwent both rs-fMRI and tb-fMRI. Time (i.e., 47 timepoints for both rs-fMRI and tb-fMRI) was added into the model as a fixed effect to analyze the overall linear trend across subjects by timepoint and is referred to as time effect. We included age, sex, and age–sex interaction into the model as covariates to mitigate bias and confounding effects. The study type (i.e., task vs. resting) was added to the models for the paired comparisons. Model fit and diagnostics were assessed per usual common practice. Linear trends over time per network within each subset were described by median and IQR (25th–75th percentile). For the group of 10 participants who underwent both rs-fMRI and tb-fMRI, the median (IQR) absolute correlation difference per overlapping network between rs-fMRI correlation and tb-fMRI correlation was calculated and reported.

A threshold of p = .05 determined change in linear trend over time. F values reflect the precise location of the point estimates on the F distribution and are additionally shown with their corresponding p values. Statistical evidence was based on the strength of the correlation coefficients, model fit, and p values. SAS v9.4 (SAS Institute, Inc.) was used for all the analyses.

Results

Participants' Demographics

Participants who underwent tb-fMRI were not demographically different from those who only underwent rs-fMRI (see Table 3). There were no differences in sex, χ²(1) = 0.31, p = .69; race, χ²(2) = 1.96, p = .47; handedness, χ²(1) = 0.75, p = .56; or years of education (Z = −0.95, p = .34) between the two groups. Participants were aged between 23 and 63 years old. The 23 participants with rs-fMRI (Mdn = 27, IQR = 23–47 years) were similar in age as the 12 participants with tb-fMRI (Mdn = 26, IQR = 23–52 years, Z = 0.31, p = .75).

Table 3.

Demographic information of participants who only underwent rs-fMRI and those who underwent both tb-fMRI and rs-fMRI.

Variable		Resting only rs-fMRI (n = 13)	Overlap rs- and tb-fMRI (n = 10)
Age (years)	Mdn (IQR)	27 (23–39)	26 (23–51)
	Range	21–63	21–63
Sex	Male	8 (62%)	5 (50%)
	Female	5 (38%)	5 (50%)
Race/ethnicity	African American	5 (38%)	3 (30%)
	Caucasian	7 (54%)	4 (40%)
	Other	1 (8%)	3 (30%)
Handedness	Left	1 (8%)	2 (20%)
	Right	12 (92%)	8 (80%)
Education (years)	< 8	0 (0%)	0 (0%)
	9–12	1 (8%)	0 (0%)
	12–15	2 (15%)	2 (20%)
	16	3 (23%)	7 (70%)
	> 16	7 (54%)	1 (10%)

Note. rs-fMRI = resting-state functional magnetic resonance imaging; tb-fMRI = task-based functional magnetic resonance imaging; IQR = interquartile range.

FC Linear Trends Over Time

Overall, there was a time effect (i.e., changes in FC linear trends by time) only in tb-fMRI (see Table 4). None of the resting-state FC network dynamics showed a time effect. Dynamic FC showed more fluctuation in tb-fMRI in inter- and intrahemispheric networks in the 12 participants (p = .002 in two common networks) than rs-fMRI in the 23 participants (p > .99).

Table 4.

Correlation characteristics and linear trends over time in resting dynamic functional connectivity networks and task-based dynamic functional connectivity networks (intrahemispheric and interhemispheric).

Hemisphere	Connections		Resting (n = 23)			Task (n = 12)
Left	ROI 1	ROI 2	Correlation, Mdn (IQR)	F	p	Correlation, Mdn (IQR)	F	p
1	Cuneus	Lingual	.80 (.73–.84)	.27	1.0	.79 (.66–.88)	1.77	.002
2	Cuneus	Pericalcarine	.76 (.69–.84)	.25	1.0	—	—	—
3	Fusiform	Lateral occipital	.69 (.56–.75)	.07	1.0	—	—	—
4	Inferior parietal	Precuneus	.62 (.52–.74)	.39	1.0	—	—	—
5	Lingual	Pericalcarine	.82 (.78–.85)	.38	1.0	.82 (.70–.90)	.84	.76
6	Pars opercularis	Supramarginal	.56 (.45–.66)	.07	1.0	—	—	—
7	Postcentral	Precentral	.80 (.70–.87)	.22	1.0	.82 (.70–.90)	.55	.99
8	Rostral middle frontal	Superior frontal	—	—	—	.68 (.51–.81)	.65	.96
Right	ROI 1	ROI 2	Correlation, Mdn (IQR)	F	p	Correlation, Mdn (IQR)	F	p
1	Cuneus	Lingual	.80 (.75–.84)	.09	1.0	—	—	—
2	Cuneus	Pericalcarine	.76 (.67–.83)	.17	1.0	—	—	—
3	Fusiform	Lateral occipital	.69 (.58–.75)	.09	1.0	—	—	—
4	Inferior parietal	Middle temporal	.62 (.52–.69)	.31	1.0	—	—	—
5	Inferior parietal	Precuneus	.68 (.59–.75)	.10	1.0	—	—	—
6	Isthmus cingulate	Precuneus	.75 (.67–.82)	.03	1.0	—	—	—
7	Lingual	Pericalcarine	.80 (.73–.85)	.16	1.0	.84 (.71–.91)		1.0
8	Pars opercularis	Pars triangularis	.67 (.57–.76)	.07	1.0	—	—	—
9	Postcentral	Precentral	.81 (.75–.86)	.09	1.0	.84 (.72–.92)		.96
Interhemispheric	ROI 1 (left)	ROI 2 (right)	Correlation, Mdn (IQR)	F	p	Correlation, Mdn (IQR)	F	p
1	Cuneus	Cuneus	.88 (.84–.91)	.11	1.0	—	—	—
2	Cuneus	Lingual	.79 (.72–.83)	.22	1.0	.78 (.66–.87)	.87	.71
3	Cuneus	Pericalcarine	.74 (.66–.82)	.18	1.0	—	—	—
4	Fusiform	Fusiform	.77 (.71–.83)	.45	.99	—	—	—
5	Inferior parietal	Inferior parietal	.63 (.53–.74)	.07	1.0	—	—	—
6	Lateral occipital	Fusiform	.65 (.52–.73)	.09	1.0	—	—	—
7	Lateral occipital	Lateral occipital	.86 (.81–.92)	.19	1.0	.86 (.76–.93)	1.75	.002
8	Lingual	Cuneus	.78 (.70–.82)	.06	1.0	—	—	—
9	Lingual	Lingual	.89 (.83–.92)	.02	1.0	.84 (.75–.93)	.42	1.0
10	Lingual	Pericalcarine	.79 (.74–.84)	.03	1.0	—	—	—
11	Pericalcarine	Cuneus	.74 (.64–.83)	.28	1.0	—	—	—
12	Pericalcarine	Lingual	.78 (.72–.84)	.14	1.0	.83 (.68–.89)	.27	1.0
13	Pericalcarine	Pericalcarine	.91 (.88–.94)	.07	1.0	.85 (.71–.90)	.74	.90
14	Postcentral	Postcentral	.86 (.75–.90)	.08	1.0	—	—	—
15	Postcentral	Precentral	.76 (.65–.83)	.02	1.0	.75 (.60–.85)	1.06	.37
16	Precentral	Postcentral	.74 (.63–.82)	.24	1.0	—	—	—
17	Precentral	Precentral	.84 (.78–.88)	.08	1.0	.83 (.73–.90)	.49	1.0
18	Precuneus	Isthmus cingulate	.73 (.65–.79)	.02	1.0	—	—	—
19	Precuneus	Precuneus	.84 (.77–.89)	.21	1.0	.83 (.71–.90)	.75	.89
20	Superior frontal	Superior frontal	.73 (.59–.79)	.06	1.0	.77 (.62–.87)	1.03	.42
21	Superior parietal	Superior parietal	.83 (.75–.91)	.02	1.0	.85 (.74–.91)	.37	1.0
22	Superior temporal	Superior temporal	.81 (.75–.85)	.02	1.0	.80 (.67–.90)	1.12	.29
23	Insula	Insula	.71 (.64–.79)	.06	1.0	—	—	—
24	Supramarginal	Supramarginal	—	—	—	.78 (.65–.88)	.77	.87

Note. Results are from models that tested for time effect and adjusted for age, sex, and age–sex interaction where applicable. IQR = interquartile range; — = common connection not found. Bold values indicate significance at p < .05.

Dynamic rs-fMRI

Sixteen common intrahemispheric networks (seven left- and nine right-hemispheric networks) and 23 common interhemispheric networks were identified in dynamic rs-fMRI. There was no time effect in dynamic rs-fMRI in 23 healthy participants (p > .99). FC linear trends did not change over time on any of the intra- and interhemispheric networks, accounting for age, sex, and age–sex interaction (see Table 4). In 10 participants who had both rs-fMRI and tb-fMRI, there also was no time effect in any of the common intra- and interhemispheric networks (see Table 5).

Table 5.

Correlation differences and linear trends over time in the subset of 10 participants with both resting dynamic functional connectivity networks and task-based dynamic functional connectivity networks (intrahemispheric and interhemispheric).

Hemisphere	Connections		Resting (n = 10)			Task (n = 10)			Resting vs task (n = 10)
Left	ROI 1	ROI 2	Correlation, Mdn (IQR)	F	p	Correlation, Mdn (IQR)	F	p	Absolute correlation difference, Mdn (IQR)	F	p
1	Cuneus	Lingual	.82 (.79–.85)	.21	1.0	.81 (.69–.90)	2.31	< .001	.10 (.05–.18)	19.49	< .001
2	Lingual	Pericalcarine	.82 (.79–.85)	.83	.78	.82 (.71–.90)	.66	.96	.08 (.04–.16)	32.38	< .001
3	Postcentral	Precentral	.82 (.73–.88)	.16	1.0	.82 (.70–.89)	.52	1.0	.10 (.04–.18)	9.52	.002
Right	ROI 1	ROI 2	Correlation, Mdn (IQR)	F	p	Correlation, Mdn (IQR)	F	p	Absolute correlation difference, Mdn (IQR)	F	p
1	Lingual	Pericalcarine	.81 (.74–.86)	.24	1.0	.84 (.71–.91)	.71	.92	.10 (.04–.18)	.01	.92
2	Postcentral	Precentral	.82 (.76–.88)	.15	1.0	.83 (.71–.92)	.85	.74	.13 (.08–.21)	2.50	.11
Interhemispheric	ROI 1 (Left)	ROI 2 (Right)	Correlation, Mdn (IQR)	F	p	Correlation, Mdn (IQR)	F	p	Absolute correlation difference, Mdn (IQR)	F	p
1	Cuneus	Lingual	.81 (.75–.85)	.17	1.0	.79 (.67–.88)	.80	.82	.10 (.04–.17)	13.45	< .001
2	Lateral occipital	Lateral occipital	.91 (.83–.93)	.22	1.0	.86 (.77–.93)	1.28	.11	.09 (.04–.16)	61.53	< .001
3	Lingual	Lingual	.91 (.89–.93)	.75	.88	.87 (.78–.94)	.65	.96	.07 (.03–.13)	141.46	< .001
4	Pericalcarine	Lingual	.81 (.75–.85)	.59	.98	.83 (.69–.89)	.29	1.0	.10 (.05–.16)	22.53	< .001
5	Pericalcarine	Pericalcarine	.93 (.88–.95)	.02	1.0	.84 (.73–.90)	.66	.96	.09 (.04–.18)	246.0	< .001
6	Postcentral	Precentral	.79 (.66–.85)	.15	1.0	.75 (.59–.85)	1.09	.32	.15 (.06–.27)	24.42	< .001
7	Precentral	Precentral	.84 (.78–.89)	.13	1.0	.83 (.73–.91)	.32	1.0	.09 (.04–.16)	14.61	< .001
8	Precuneus	Precuneus	.89 (.82–.91)	.54	.99	.83 (.73–.91)	1.29	.10	.10 (.05–.18)	80.53	< .001
9	Superior frontal	Superior frontal	.76 (.62–.85)	.03	1.0	.75 (.59–.86)	1.08	.35	.12 (.06–.25)	.40	.53
10	Superior parietal	Superior parietal	.89 (.80–.93)	.46	1.0	.85 (.75–.92)	.36	1.0	.08 (.03–.17)	35.82	< .001
11	Superior temporal	Superior temporal	.81 (.76–.85)	.13	1.0	.78 (.66–.88)	1.22	.17	.13 (.06–.23)	62.36	< .001

Note. Results are from models that tested for time effect and adjusted for age, sex, and age–sex interaction where applicable. Bold values indicate significance at p < .05. ROI = region of interest; IQR = interquartile range.

Dynamic tb-fMRI

Overall, there were fewer common networks in tb-fMRI than rs-fMRI, with six common intrahemispheric networks (four left- and two right-hemispheric networks) and 12 common interhemispheric networks identified using dynamic tb-fMRI (see Table 3). In 12 participants who underwent tb-fMRI, changes in linear trend over time (accounting for age, sex, and age–sex interaction) were observed in one common network in the left hemisphere (left cuneus–left lingual connection, F = 1.77, p = .002), and one interhemispheric connection (left lateral occipital–right lateral occipital connection, F = 1.75, p = .002; see Figure 4). Linear trend was stable over time in these two networks in the right hemisphere (p > .96).

Figure 4.

Results from repeated-measures mixed models that examined correlation fluctuations and linear trends over time in rs-fMRI (n = 23) and tb-fMRI (n = 12). There were two connections that showed linear trends over time (one common network in the left hemisphere and one interhemispheric network): (a) Correlation over time per subject in the left cuneus–left lingual intrahemispheric connection and left lateral occipital–right lateral occipital interhemispheric connection in rs-fMRI (n = 23). (b) Correlation over time per subject in the same connections in tb-fMRI (n = 12). Each dotted line represents correlation linear trend over time in each participant. rs-fMRI = resting-state functional magnetic resonance imaging; tb-fMRI = task-based functional magnetic resonance imaging.

Similarly in the 10 participants who had both rs-fMRI and tb-fMRI, changes in linear trend over time were observed in the same common network in the left hemisphere (left cuneus–left lingual connection, F = 2.31, p < .001; see Table 4), accounting for age, sex, and age–sex interaction. Dynamic FC network linear trend over time was markedly different between rs-fMRI and tb-fMRI among 10 participants who underwent both rs-fMRI and tb-fMRI, accounting for age, sex, and age–sex interaction (see Table 4). More specifically, there was a difference in FC network linear trend over time on all common networks in the left hemisphere (i.e., left cuneus–left lingual connection, F = 19.49, p < .001; left lingual–left pericalcarine connection, F = 32.38, p < .001; and left postcentral–left precentral connection, F = 9.52, p = .002) in tb-fMRI (see Figure 5). There also was a difference in FC linear trends over time in all except one interhemispheric network (i.e., left superior temporal–right superior temporal connection) in tb-fMRI (p < .001 for 10 networks), with no differences between tb-fMRI and rs-fMRI linear trends over time in the FC networks activated in the right hemisphere (p > .11).

Figure 5.

Results from repeated-measures mixed models that examined differences in correlation fluctuation and linear trends over time between rs-fMRI and tb-fMRI among participants who underwent both rs-fMRI and tb-fMRI (n = 10). There were 13 connections that showed linear trends over time (all common networks in the left hemisphere and 10 interhemispheric networks). (a) Correlation over time per subject in the left-hemispheric connections. (b) Correlation over time per subject in the interhemispheric connections in participants who underwent both rs-fMRI and tb-fMRI (n = 10). The top 10 figures show the connections in rs-fMRI. The bottom 10 figures show the same connections in tb-fMRI. Each dotted line represents linear trends over time in each participant. rs-fMRI = resting-state functional magnetic resonance imaging; tb-fMRI = task-based functional magnetic resonance imaging.

Discussion

We investigated the dynamics of FC during an auditorily presented language comprehension task in rs-fMRI and tb-fMRI in healthy participants. Dynamic FC linear trend over time was examined in the intrahemispheric (left and right hemispheres) and interhemispheric networks.

Network Correlation Dynamics Between rs-fMRI and tb-fMRI

Importantly, our results demonstrated a difference in FC network dynamics between rs-fMRI and tb-fMRI for all activated networks in the left and intrahemisphere but for none of the FC networks activated in the right hemisphere. None of the intra- and interhemispheric connections in rs-fMRI showed changes in linear trends over time during auditory comprehension.

These results support our hypothesis of differential FC network dynamics in tb-fMRI and do not support previous research studies that have advocated for the interchangeability of using static rs-fMRI and tb-fMRI in determining FC in language-related networks (Rolinski et al., 2020). Poeppel and colleagues suggest that speech and language information is processed by networks using multiple timescales (Arnal et al., 2015; Poeppel, 2003; Teng et al., 2016, 2017, 2019). For instance, phonetic, syllabic, and intonation information is conveyed at different timescales. As such, during auditory comprehension, the auditory system integrates the multiscale information over time from continuously varying acoustic signals for higher level language processing (Liegeois-Chauvel, 2004; Peelle & Davis, 2012; Poeppel & Assaneo, 2020; Rimmele et al., 2023). The findings of the current study support the multiscale processing model (Giraud & Poeppel, 2012; Wu et al., 2021). It is known that FC in tb-fMRI reflects task-related neural networks including cognitive processes (Betzel et al., 2020; Hutchison et al., 2013; Kucyi & Davis, 2014; Kurtin et al., 2023; Song & Lee, 2023). Our analyses indicated that the same neuronal networks analyzed in the same participants demonstrated increased dynamic FC fluctuation during language processing in tb-fMRI than the same network in rs-fMRI. We think that this partly reflects the multiple timescales at which different aspects of language processing are being temporally encoded.

Poeppel (2003) further proposes that although both left and right hemispheres participate in speech processing, the left hemisphere responds better to faster modulated signals, while the right hemisphere responds better to slower signals (Arnal et al., 2015; Poeppel, 2003). Although the current study did not intend to test for cerebral hemispheric differences, we found, using an entirely whole-brain computational approach, that more changes in FC linear trend over time were identified in specific left-hemispheric connectivity networks but in none of the common networks identified in the right hemisphere. This finding also supports Poeppel's multiscale processing model. Simonyan et al. (2013) also reported dorsal striatal endogenous dopaminergic release subserving FC changes in the left hemisphere during sentence production in healthy individuals (Simonyan et al., 2013). They indicated that speech-induced dopamine release was significantly associated with neural activity that influenced the lateralization of FC results. Similarly, using an entirely noninvasive neuroimaging approach of language comprehension, our findings also raised the possibility of using network connectivity dynamics to identify language-specific networks and to separate such networks from those that are processing different, less varied and non-multiscaled information, such as attention, in patients who may require such determination prior to clinical interventions such a fractionated focal radiation therapy (e.g., Gamma Knife or proton therapy) and/or presurgical intervention for improved patient outcomes. For this reason, future research will have to evaluate its sensitivity in consistently determining language-related versus language-specific real-time determination in comparison to ROI fMRI neuroactivation studies and/or intra-operatively determined language laterality, the current gold standard.

The results of the current study partially corresponded with Elton and Gao (2015), who indicated a significant decrease in correlation fluctuations in some of the networks during a visual classification task in tb-fMRI compared to rs-fMRI in 19 healthy participants (Elton & Gao, 2015). However, in the same study, they also indicated a significant increase in correlation variability in some other networks during the same task. On the other hand, the results of the current study did not support Chen et al. (2015), who reported reduced correlation variability during a 2-back working memory task in tb-fMRI compared to rs-fMRI in 21 healthy participants (Chen et al., 2015). However, this apparent discrepancy may indicate that the network dynamics identified in our study reflect patterns that are language specific rather than general cognitive encoding processes. In addition, differences in network correlation analysis methods may also have contributed to these differences.

An alternative explanation for the difference in network correlation characteristics between rs-fMRI and tb-fMRI is mixed cerebral dominance for language processing. In the current study, 12% of the participants were left-handed; however, there was no statistical difference in handedness between participants who underwent both rs-fMRI and tb-fMRI. Thus, we think that this is an unlikely explanation. Moreover, differences in FC linear trends over time were noted only in the left- and interhemispheric networks, but not in any of the right-hemispheric networks.

Clinical Implications

Clinically, the results of the current study imply that the dynamics of language-specific network correlations may be an important element in understanding a patient's complete neurolinguistic profile. Current neurocomputational models of language processing predominantly employ static rs-fMRI. However, temporal FC dynamics during language processing are not well captured via static connectivity network correlations. It is highly important to consider the dynamics of network correlations for developing precise neurocomputational models with FC patterns for the prediction of language functions following CNS injury and for identifying language recovery trajectories.

Moreover, in participants who underwent both rs-fMRI and tb-fMRI, specific network dynamics in the left hemisphere showed changes over time during tb-fMRI, while no statistically significant network correlation linear trend over time was seen for any of the common networks identified in the right hemisphere. These differences suggest that the left-hemispheric common network dynamics in tb-fMRI may become potential biomarkers for language processing, language laterality, and possibly language recovery trajectories, although further longitudinal research in patients with language disorder is needed. It is difficult to predict language recovery in patients with CNS injury, including people with aphasia, because the degree and type of language deficits are heterogeneous (Brady et al., 2016; Sheppard & Sebastian, 2021). It is crucial to identify robust predictors for language recovery in this population in order to provide evidence-based personalized speech and language therapy in order to maximize their potential recovery from language deficits (Kristinsson et al., 2022). Temporal dynamics of the left-hemispheric network correlations may significantly contribute to the prediction of language recovery in patients with CNS injury. Thus, neurocomputational modeling of dynamic FC, along with traditional language evaluations, may further improve individual outcome predictions. Future research will be needed to further investigate this possibility as well as its contribution to individualized rates of recovery.

Limitations

When considering the results of this study, several caveats should be noted. First, we employed a conservative statistical approach in order to ensure FC network robustness and reliability. We did so by including only network correlations that were identified in all the participants. Therefore, neural networks may be censored if not statistically significant in one subject, and we may have missed other relevant networks.

The age effect was statistically controlled for in the main analyses. While most participants were young (median age = 27 years), there was a participant who was 63 years old. This skewed distribution may have contributed to the age effect.

Participants performed the ADDT during tb-fMRI. The networks identified in the current study may not reflect other nonauditory comprehension tasks.

Although we conducted a statistical power calculation and analyzed a total of 59,831 data points (42,159 in rs-fMRI, 10,152 in tb-fMRI, and 7,520 in overlap), the number of subjects is relatively small. Hence, our results may only reflect the performance of this cohort and may have inflated intersubject correlation differences. For these reasons, future research using a larger cohort of subjects with a lower participation threshold may be needed to determine the generalizability of these findings.

Conclusions

The current study retrospectively investigated the stability of language neural networks for dynamic FC in rs-fMRI and tb-fMRI during auditory comprehension in healthy adults. The findings of the current study revealed that the temporal dynamics of identical neural networks during auditory comprehension demonstrated increased linear trends over time in specific left-hemisphere networks during tb-fMRI compared to the resting-state network dynamics. In contrast, these dynamic FC changes were not observed in the right-hemisphere networks. We think that these findings may be an important contributing factor in determining language-related clinical profiles of patients following CNS injury as well as in developing computational models that can predict language-related functional neural recovery. Further studies with larger cohorts of healthy participants as well as patients with language disorders are warranted to confirm the generalizability of the results of the current study in predicting functional neural recovery.

Data Availability Statement

Upon reasonable request, the de-identified data used in the current study are available from the National Institutes of Health Biomedical Translational Research Information System with permission from our co-authors, Sara Inati (sara.inati@nih.gov) and William Theodore (william.theodore@nih.gov).

Funding Statement

The study was funded by the National Institutes of Health (NIH) Intramural Program, the National Institute on Deafness and Other Communication Disorders, the National Institute of Neurological Disorders and Stroke, and the NIH Clinical Center (RASCL Z01, award number ZIA CL090077, awarded to Nadia Biassou).