This literature review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (29), as depicted in Figure 1. PubMed, Web of Science, Google Scholar, and ScienceDirect were queried with keywords to generate articles. Articles prior to July 2020 were included, yielding 1,097 records (database search in Figure 1). Additional targeted searches were conducted to identify articles from known rsfMRI researchers prior to July 2020 that were not captured by the four database searches, yielding eight records (targeted author search in Figure 1). In total, we identified 1,105 records for screening.

Keywords for PubMed included: (traumatic brain injury [Title/Abstract]) OR (tbi [Title/Abstract]) AND (rsfmri [Title/Abstract]) OR (resting-state [Title/Abstract]). Keywords for Web of Science included: TOPIC: (“resting state fMRI” “traumatic brain injury”) OR (“rsfMRI” “TBI”) OR (“resting state” “MRI” “traumatic brain injury”) OR (“rsfMRI” “traumatic brain injury”) OR (“resting state fMRI” “TBI”) across all years and databases. Keywords for Google Scholar included: “traumatic brain injury” “resting state functional MRI” -rats -dementia -"hepatic encephalopathy” - “task based.” Keywords for ScienceDirect included: Title, abstract, keywords: (“Traumatic brain injury” or “TBI”) AND (“resting state” OR “rs-fmri”).

Exclusion criteria included: non-human subjects, no TBI subjects, no matched controls, no rsfMRI performed, task-based fMRI only, abstracts, dissertations, case studies, case series, literature reviews, studies that only performed analyses using graph theory, machine learning, other novel methods, or did not report within-network FC results (Figure 1).

Query results were converted to the .bib format and imported into a shared folder in Mendeley. Two rounds of selection were performed by evaluating titles and abstracts. Among six reviewers, two independently reviewed each manuscript for inclusion/exclusion criteria. If they disagreed, Dr. Bickart made the final decision. After screening, 910 manuscripts were excluded, 75 of which were duplicates, leaving 195 papers for final eligibility screening (Figure 1).

Next, the reviewers read each paper completely and excluded an additional 145 papers based on the same inclusion/exclusion criteria as above (Figure 1). In total, three papers via targeted search and 47 papers via database search were included, yielding 50 papers.

The following information was extracted from each manuscript: mechanism of injury, TBI severity, TBI sample information (age statistics, gender, sample size), control sample information, time post-injury, cross-sectional and/or longitudinal results, scanner manufacturer, tesla strength, coil channel, scan duration, slice thickness, eyes open or closed, sequences acquired during fMRI acquisition, image processing pipeline, software used, rsfMRI metrics computed, analysis type, functional networks/regions investigated, and statistics reported. For longitudinal studies, we extracted the results of cross-sectional, between-group FC comparisons at different time points (i.e., acute, subacute, or chronic), but did not extract within-group FC comparisons over time.

For comparison purposes, the data extracted from each paper was organized into seven canonical networks based on a widely used and previously defined parcellation scheme (30, 31). The networks included Default Mode (DMN), Executive Control (ECN), Limbic (LN), Dorsal Attention (DAN), Salience (SN) or Ventral Attention (VAN), Visual (VN), and Sensorimotor (SMN) (Figure 2). Findings reported for seed-based or region-of-interest (ROI) analyses were assigned to the network that incorporated that seed or ROI according to the map depicted in Figure 2 and region assignment listed in Table 1. Most studies that outlined their own seed/ROI network assignments adhered to this model. With the data organized into these networks, we could then tabulate results on a comparison-by-comparison basis. Given that studies ranged in the number of networks they analyzed, we report findings for every comparison made across the 50 studies included (Table 2). We only extracted and reported results from the within-network analyses and excluded graph, between-network, and other analyses to enable a semiquantitative summary of results—quantifying the proportion of comparisons per network that were significantly different between TBI and control populations.

The original authors of all included papers were emailed to verify our interpretation of the results for this review and to obtain additional statistics for a meta-analysis. Nineteen of the lead authors were able to both verify our interpretation and provide the relevant statistics. This was, however, not enough to perform a meta-analysis.

After the data was validated, the number of reported significant increases, decreases, and no differences in connectivity were manually counted by the reviewers. To ensure accurate transmission of article annotation, code was written to count and organize the data into a database (Literature Review Toolkit: rsfMRI in TBI [Source code]).¹ The manual counts and code were compared and mistakes were corrected until both were in concordance.

To determine if findings varied by injury characteristics, demographic variables, or methodological parameters, we parsed the findings into categories: injury severity, age, chronicity, population type, and a variety of imaging parameters (Figure 3). These categorizations are derived from the prevailing consensus across TBI literature.

For severity, we classified results pertaining to TBI patients based on their Glasgow Coma Scale (GCS) scores (32, 33). GCS scores greater than or equal to 13 were classified as mild, GCS scores ranging from 9 and 12 were considered moderate, and GCS scores less than or equal to 8 were considered severe. Some studies included TBI participants of varying severities. Therefore, we also categorized findings as mild/moderate, moderate/severe, or mixed (i.e., mild, moderate, and severe for such cases).

For age, we classified studies involving subjects aged 12–18 as adolescent, age 18 and greater as adult, and otherwise as mixed. Studies that met the inclusion criteria did not identify participants younger than 12 years old.

For chronicity, there is no standard or agreed-upon definition of acute, subacute, and chronic. We chose to classify TBI chronicity based on sources that use biomarkers to track injury progression. For example, such studies (34, 35) define acute TBI as within days, subacute within weeks, and chronic as months to years post-injury. Clinically, there is also some consensus around chronicity as it relates to post-concussion syndrome (PCS). Specifically, PCS refers to a characteristic cluster of symptoms lasting greater than or equal to 3 months after injury (36). We therefore classified scans acquired within the first week post-injury as acute, between 1 and 12 weeks post-injury as subacute, and >12 weeks post-injury injury as chronic. Again, some studies included TBI participants of varying chronicities. Therefore, we also categorized findings as acute/subacute and subacute/chronic to include these comparisons. For longitudinal studies with comparisons between TBI and matched control groups at various times since injury, we treated each comparison as unique, rather than reporting changes in FC over time for either group individually.

For population type, we categorized studies by sport (Sport), military personnel (Military), civilians not injured in sport (Civilian), and mixed if participants were from an array of the aforementioned populations (Mixed).

We pooled and parsed the findings from comparisons of TBI patients to matched controls, counting the number of reported significant increases and decreases in connectivity or no significant differences in connectivity. Significance was determined by the authors of each study. We would have preferred to perform a meta-analysis to account for differences in effect sizes across studies, but very few studies reported or responded to our inquiries to produce the necessary statistics for this. Comparisons were first collapsed across injury characteristics, such that the connectivity differences could be compared across each of the canonical networks (Figure 4). Next, comparisons were collapsed across networks, allowing for connectivity differences to be compared across injury characteristics (Figure 5). Lastly, connectivity differences were examined across both network and injury characteristics to allow for analysis of each TBI category or demographic within a specified network (Figure 6).

Alongside analysis collapsed by injury characteristics or network, percent global signal was calculated. Percent global signal was defined as the number of increases and decreases in connectivity relative to matched controls divided by the total number of connectivity comparisons. The goal was to use percent signal as a proxy for the degree to which networks reported any difference in connectivity following TBI compared to controls, regardless of whether the connectivity increased or decreased (Figures 4, 5).

From 1,105 manuscripts, 50 met inclusion criteria, reporting a total of 179 network connectivity comparisons between a total of 1,397 TBI patients and 1,179 matched controls; an average of 3.58 comparisons per manuscript. TBI cohorts primarily included civilian adults with a mild severity of TBI in the subacute to chronic stage (Figure 3).

Of the 179 comparisons, the majority were conducted in DMN (44%) and the least in the DAN (4%) (Figure 4b). Collapsing across injury characteristics and networks, 25% of comparisons showed increased connectivity, 22% showed decreased connectivity, and 53% showed no significant differences between the TBI and control groups, yielding a percent signal of 47% (Figures 4a–c).

When collapsing across all networks, injury characteristics, patient demographics, and imaging techniques (Figure 4b, “Total” row), the findings were generally mixed. There were about an equal number of comparisons showing a significant difference versus no difference in connectivity (47% signal) between patients with TBI and matched controls. Of the significant differences, again, the findings were fairly mixed, but with slight predominance of increased connectivity (45 increased and 39 decreased).

Parsing the comparisons by severity revealed that patients with severe and moderate/severe TBI were more likely to show increased (67%) over decreased connectivity (33%) as compared to controls, whereas patients with mild TBI showed a 50/50 split between increased and decreased connectivity as compared to controls (Figure 5). It may be that higher severity injuries disrupt more axonal connections, warranting compensatory processes (37–39), such as cortical reorganization, neurogenesis, axonal sprouting, and angiogenesis, which have been demonstrated in animal models (40, 41).

Surprisingly, however, studies show a similar likelihood of finding any significant difference in FC whether in mild TBI (52%) or moderate/severe TBI combined (56%) as compared to controls (Figure 5). Despite the obvious difference in structural and neurological consequences of injury between these TBI severities, the differences in FC as compared to controls is not as apparent. Studies directly comparing FC in patients across TBI severities are lacking.

Parsing the comparisons by chronicity revealed a cross-sectional trend of increased connectivity by time since injury. Because we dealt with longitudinal studies as discrete, between-group comparisons of FC at acute, subacute, or chronic times since injury, our descriptive findings do not enable conclusions about within group changes in FC over time. Nevertheless, cross-sectionally, patients with TBI who were imaged in the acute and acute/subacute window showed less increased (42%) than decreased connectivity (58%) as compared to controls (Figure 5). From the subacute to chronic window however, the differences flipped to more increased (60%) than decreased connectivity (40%) as compared to controls. Though these results are cross-sectional, this trend suggests connectivity may increase over time post-injury.

Several of the studies included in our review contained longitudinal data, the majority of which report a relative within-group increase in connectivity at the chronic as compared to the more acute time points post-injury (21, 42, 43). Studies that did not find increased connectivity in the chronic time points showed no difference as a function of time since injury (44, 45) or revealed more complex findings, such as increased and decreased connectivity within a single network as a function of time since injury (e.g., increased connectivity at the left anterior prefrontal cortex, but decreased connectivity at the middle cingulate gyrus) (46). Most notably, however, no longitudinal study reported a whole-network decrease in connectivity in the chronic stage. Studies utilizing graph theory have also noted a pattern of initially decreased to subsequently increased connectivity over time since injury (15, 47, 48). Increased connectivity has correlated with better function (42) and worse function (49) demonstrating again that the direction of connectivity change or difference compared to control may not be as important as the finding of any change or difference. These longitudinal effects have been interpreted as adaptive or compensatory when relating to improved function (44, 50, 51).

Parsing the comparisons by age did not reveal any discernible pattern. Trends based on age alone (adolescent vs. adult) are difficult to draw given the large majority (90%) of studies included only adults. Furthermore, adult age ranges were restricted, with the first quartile less than 26.88 years and the fourth quartile greater than 38.49 years. Because most of the TBI patients were scanned in early adulthood, we cannot infer whether connectivity patterns emerge after injury in younger and older ages due to the paucity of studies overall, and in our study in particular, for these age groups.

Parsing comparisons by population type revealed a notably larger percent signal (71%) in the military population. Overall, however, the military population showed an even split of increased and decreased connectivity findings, which is consistent with the other populations studied. The military population was also the least studied (7.8% of total comparisons). Because injury in military personnel can include unique mechanisms and are often comorbid with sequelae of psychological trauma as compared to civilian and sport-related injuries (52), it is possible that the brain differences are more striking.

The DMN was the most studied network (78 comparisons) and had the highest signal of any of the networks included (55% signal). The ratio of increased to decreased connectivity between people with TBI and matched controls was, however, roughly equal (23–20, respectively). It is well known that the DMN plays a critical role in intrinsic cognitive processes. Though the DMN generally had mixed results across most patient characteristics, it trended toward increased connectivity in the moderate/severe TBI population and at the subacute time frame (Figure 6b, column 1). Previous studies have interpreted this increased connectivity as a compensatory or adaptive mechanism because of its association with positive cognitive outcomes (44, 50, 51). It is possible compensatory processes do not readily occur unless the injury is sufficiently damaging. Perhaps such compensation does not occur until at least weeks to months after the injury, which may follow periods of increased metabolic activity (53, 54).

The DAN showed the lowest signal (14%) of all the networks, suggesting this network may not be affected by TBI and/or whatever changes that may occur are not sensitive to rsfMRI methods. Of note, the DAN also had the lowest number of total comparisons across the literature (Figure 4). Further investigation of the DAN may be required to rule out the DAN as a potential marker of TBI. Alternatively, this may be a network to leave out of analyses to retain valuable degrees of freedom.

The SN had the next lowest percent signal (39%) (Figure 4). Given the putative role of the SN in affective processes (55, 56), including affective dysregulation in some patients with TBI (57), its low signal is surprising. However, because the SN serves as a switch between intrinsic and extrinsic cognitive processes (58), rsfMRI may not be sensitive enough to capture changes in connectivity without patients performing goal-directed activity or without looking at internetwork connectivity differences. Again, further investigation of this network at rest and during tasks will provide greater insight to its sensitivity as a potential biomarker of TBI.

Parsing the comparisons by whether patients were instructed to keep their eyes open, fixated, or closed revealed a striking pattern. Eyes open or fixated showed 75 and 64% signal, respectively, whereas eyes closed only revealed 33% signal (Figure 7). Interestingly, about half of the studies instructed participants to close their eyes, despite newer evidence that eyes open may enhance connectivity strengths (59).

Generating quartiles of the number of volumes captured during scans revealed a notably higher percent signal among the third quartile (73%), which had a range of 182–245 volumes per scan (Figure 7). In contrast, all other quartiles had a markedly lower percent signal (<48%). These results suggest the number of volumes captured during scans may have an ideal value that could serve to increase the sensitivity of rsfMRI in detecting connectivity differences.

Generating quartiles of the scan duration revealed a notably lower percent signal among the second quartile (37%), which had a range of 315–360 s (Figure 7). In contrast, the first, third, and fourth quartiles all demonstrated markedly higher percent signals. Recent studies support greater reliability with increasing scan length, though that cutoff point is debated and not well defined (60, 61).

Other than GE, which was rarely used, scanner manufacturer and coil channel did not meaningfully impact the findings in that there were similar proportions of increased, decreased, and null findings no matter which scanner or head coil was used (Figure 7). GE did demonstrate a percent signal of 77% suggesting more likely to show some difference across groups as compared to null but only four studies utilized GE scanners, which is not enough to make a strong conclusion. However, the fact that the other two most commonly used scanner manufacturers did not appreciably impact the findings suggests that this pattern of findings are likely generalizable and could be expected to hold across sites in future research.

Creating tertiles of the slice thickness demonstrated a substantially higher percent signal among the third tertile, which had a thickness of >4 mm per slice (Figure 7). The choice in slice thickness is determined by the time to acquire a set brain volume, spatial resolution, minimization of signal dropout, and BOLD signal-to-noise (SNR) ratio, which largely determines the statistical power to detect nodal activation. Therefore, any increase in spatial resolution with thinner slices will be offset by a reduction in the statistical significance due to increased SNR (62–64). Prior work demonstrates that thicker slices will yield more low-percentage signal changes, consistent with our results (63). However, the optimal slice thickness depends on the study, as certain brain regions may require higher resolution, which may come at the cost of sensitivity per unit time, to enhance localization and reduce signal dropout.

When evaluating preprocessing software as a factor, custom pipelines achieved a notably higher signal than the other softwares (Figure 8). It may be that custom pipelines are designed for the specific dataset obtained, and thus software can be designed to achieve higher sensitivity in detecting connectivity differences. However, because this study did not have access to the custom softwares, there is the possibility that a higher signal could result from poorly implemented pipelines that are not as rigorous at mitigating specious findings.

Parsing the comparisons by regression type revealed performing regression in all categories (global, white matter, CSF), except motion, increased the likelihood of detecting a significant difference between TBI and control populations (Figure 8). Because percent signal was similar between comparisons that did and did not perform motion regression, it appears that motion regression does not significantly affect sensitivity of finding a difference in connectivity in these studies.