Which voxel-wise resting state fMRI metric is the most discriminatory for concussion?  A secondary analysis.

doi:10.21203/rs.3.rs-4578572/v1

Download PDF

Research Article

Which voxel-wise resting state fMRI metric is the most discriminatory for concussion? A secondary analysis.

https://doi.org/10.21203/rs.3.rs-4578572/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Background

Concussions are among the most common sport-related injuries. While symptoms remain the most widely studied outcome, other measures such as resting state functional magnetic resonance imaging (rsfMRI) are being increasingly studied to better understand the neurophysiology of concussion. The basis for rsfMRI is the temporal blood oxygen level dependent (BOLD) signal, which shows complex changes in brain activity over the course of a scan. The BOLD signal can be operationalized in many ways, which effects how the data are analyzed and interpreted. It is therefore important for neuroimaging researchers to understand which rsfMRI metric is most discriminatory between concussion subjects and healthy controls, as such knowledge may aid in the design of future studies. The primary purpose of our study was to employ a random forest approach to determine which BOLD signal metric (of which we selected six) was most discriminatory between concussion subjects and healthy controls. As a secondary objective, we aimed to understand which anatomical regions of interest in the brain were most discriminatory between these two groups.

Methods

We performed a secondary data analysis of prospectively collected concussion data (n = 28), alongside healthy control data retrieved through an open-source neuroimaging repository (n = 379). After pre-processing data in CONN 21a, six BOLD metrics were computed: mean, standard deviation, Lyapunov exponent, sample entropy, amplitude of low frequency fluctuations (ALFF), and fractional ALFF (fALFF). Using a 60/20/20 (training/testing/validation) split of the data, random forest models were built using 'scikit-learn' and 'imbalanced-learn' to determine which metrics and anatomical regions of interest were most discriminatory between concussion and healthy subjects.

Results

We found that ALFF was the most discriminatory BOLD metric, with an ROC AUC = 0.993. With respect to regions of interest, we found that the BOLD signal from the cerebellum, vermis, and putamen (left) were most often discriminatory.

Conclusions

Our study provides concussion neuroimaging researchers with insight into which rsfMRI metrics may be discriminatory between groups, as well as which anatomical regions may help distinguish between concussion and healthy subjects. This may inform the design of new research studies that take aim at assessing or monitoring concussion using rsfMRI.

Trial registration:

Not applicable.

Concussion

brain injury

MRI

functional MRI

resting state MRI

random forest

machine learning

feature selection

Sport-related concussions are a public health concern, with recent studies suggesting that the incidence is rising in both children and adults (1). While the majority of pediatric and adult concussion patients can expect to achieve symptom recovery within one-month of injury (2), 20% or greater may have symptoms that persist beyond this timepoint (3). However, even in cases in which symptoms resolve within expected timelines, systematic reviews suggest that symptom recovery does not equate to physiological recovery from concussion (4), as impairments can still be observed using magnetic resonance imaging (MRI), electrophysiology, autonomic markers (such as heart rate variability) in those who are asymptomatic. As such, there is a need to study concussion using outcomes other than symptoms to improve our understanding of the neurophysiology of this injury.

One of the most promising non-symptom-based outcome measures in concussion is functional MRI (fMRI) (5, 6). By definition, concussion is associated with functional impairment and not gross structural damage (2), thereby making fMRI an optimal method for studying the neuropathology of this injury. Systematic reviews have identified fMRI impairments in multiple concussion populations, including youth (7–10), sport-related concussion (11, 12), and post-concussion syndrome (13, 14). While fMRI has led to novel insight into concussion, it is an outcome measure that can be operationalized in many ways. More specifically, at a high-level categorization, fMRI can be used in a task-based manner or a task-independent manner (also known as resting-state fMRI; rsfMRI), with the majority of recent concussion research focused on the latter. Further, within the more widely researched category of rsfMRI, there are an increasing number of ways in which the blood-oxygen-level-dependent (BOLD) signal, which is the basis for fMRI, has been operationalized at a voxel or regional level (15, 16). Such differences can have an ultimate impact on the results and interpretation of a rsfMRI analysis.

Considerable research on rsfMRI in concussion has focused on the multitude of resting state networks (17). These macroscale networks (i.e. covering a large portion of the brain) are based on high temporal correlation between various brain regions (18) While such network neuroscience is an important tool in concussion and other neurological populations, voxel-wise analyses remain powerful for understanding region-specific neurophysiology (19). At a basic level, such voxel-wise metrics include the mean or variability of the BOLD timeseries can be used as an rsfMRI metric (20). More advanced measures include those which compute the of the spontaneity of the BOLD signal, to understand its fractal dimension and complexity (21, 22), which has now been applied in concussion (23). Specific frequency bands (presumed to reflect more physiologically-relevant neurological activity) can also be examined to derive metrics such as amplitude of low frequency fluctuations (ALFF) and fractional ALFF (fALFF); these metrics too have been applied in concussion (24–26). Therefore, as the number of rsfMRI metrics continues to grow, metric choice becomes a more important consideration. This is particularly notable as brain network characteristics (across different areas of neuroscience) are impacted by both metric choice and frequency band selection (27–29). The fMRI signal also varies by region in a manner dependent on physiological effects, such as cerebral blood flow and cerebrovascular reactivity (30), suggesting that there may be regional considerations in metric selection

It is important to improve our understanding of which rsfMRI metrics offer the most diagnostic utility in concussion research. To this end, we performed a study in which six rsfMRI metrics were calculated (at a regional level) in concussion patients and healthy controls. Using a random forest approach, we then determined which of these six metrics served as a better classifier between concussion and healthy neuroimaging data. We also examined whether there were certain anatomical regions of interest that were more discriminatory between groups. Our study may help inform neuroimaging concussion researchers about which metrics and regions are more discriminatory at a group level, which may aid in establishing new research questions or analysis pipelines.

Design

The pediatric concussion data were collected as part of a single-center, prospective cohort study on which the senior author was one of the lead investigators. The control data were retrieved through an open-source neuroimaging database, namely the Autism Brain Imaging Data Exchange II (ABIDE-II) repository. Open-source control data were used to increase our sample size. This study (and in particular the parent, prospective study on concussion) was approved by the Hamilton Integrated Research Ethics Board (HiREB).

Details on the following sections are reported in our prior publication on sex-effects in resting state brain activity in pediatric concussion (31). However, a brief description of each section is provided below, in addition to a more fulsome analysis discussion (which departs from the manuscript cited above).

Participants

The pediatric concussion participants (aged 9-17, n = 28) were prospectively recruited by our clinical team at McMaster University and affiliated sites, such as McMaster Children’s Hospital. All participants received a diagnosis of concussion by a clinical team member with expertise in the area, and they were (on average) one-month post-injury at the time of scanning. Most children had a sport-related concussion. Exclusion criteria included: more severe brain injury, multiple-systems injuries involving the head, and prior neurological history.

The healthy control data were obtained through ABIDE-II (n = 379), which contains >500 healthy control rsfMRI scans collected across 19 sites using 3.0T scanners (32). All possible pediatric rsfMRI scans that matched the age range of our clinical group were considered for inclusion. The scans included in our analysis were those which met quality assurance criteria (with respect to signal-to-noise ratio, percentage of outlier scans, and data smoothness). Further, we only included a scan from ABIDE-II if it included ≥180 timepoints.

MRI acquisition

Concussion subjects were all scanned at the Imaging Research Centre at St. Joseph’s Healthcare, Hamilton. All participants were scanned using the same hardware, namely a 3.0T GE Discovery MR750 scanner and a 32-channel phased array receive coil. The data acquisition (detailed in full in our prior report (31)), but in brief involved: a 3-plane localizer; 3D inversion recovery-prepped SPGR T1-weighted sequence TR/TE = 11.36/4.25 ms, TI = 450mms, flip angle = 12°, 512 × 256 matrix interpolated to 512 × 512, 22 cm axial field of view, 1 mm thick); rs-fMRI (gradient echo EPI, TR/TE = 2000/35 ms, flip angle = 90°, 64 × 64 matrix, 180 time points, 3 mm thick, 22 cm field of view). For all concussion subjects, a B₀ map (in Hz) was acquired (following the geometric prescription of the rsfMRI) using a tool available on the scanner. The clinical data were B₀ fieldmap corrected using the `epiunwap` package (33). The control data were downloaded (assuming they met the quality assurance checks, as above) and entered into our pre-processing pipeline, as per the following section. The control data did not include B₀ fieldmaps, and we were therefore unable to perform such correction on those. However, analyses from our lab suggests that while B₀correction is ideal, lack thereof may not compromise the integrity of analyses using large open-source neuroimaging samples (34).

rsfMRI pre-processing

All pre-processing was performed in CONN 21a, which uses functionality of SPM12 and MATLAB. The pre-processing pipeline included the following steps: functional realignment and co-registration to the anatomical data, warping into MNI space (using 1mm and 2mm isotropic voxels for anatomical and functional data, respectively), slice-timing correction, outlier detection using SPM’s Artifact Detection Tool, segmentation and normalization, spatial smoothing (8 mm FWHM). De-noising was then performed, using CONN’s anatomical component-based noise correction procedure (aCompCorr). This involved projecting out noise associated with subject motion, outlier scans, white matter, cerebrospinal fluid. Temporal filtering was then performed to remove frequencies outside of the 0.008-0.1Hz range. Cortical and sub-cortical anatomical regions of interest (ROIs) were extracted using the Harvard-Oxford atlas, included in the CONN installation, which identifies 164 anatomical ROIs.

The BOLD signal for each subject and each region of interest (ROI) was extracted from CONN after pre-processing. All metrics of interest (mean, standard deviation, sample entropy, Lyapunov exponent, ALFF, fALFF; see Table 1 for more details) were computed in MATLAB on the extracted BOLD signals. More specifically, mean, standard deviation, and Lyapunov exponent for each time series were calculated using MATLAB’s mean(), std(), and Lyapunov Exponent with fs = 0.5Hz (i.e. the sampling frequency of the data) functions, respectively. A sample entropy algorithm (35,36) was written in MATLAB and applied to the BOLD signals with dim = 3, and r = 0.6 (values taken from (35) as that work was optimized for rsfMRI analysis, where dim is the rasterizing window size and r is the maximum threshold entropic distance a data point in the window can be from other data points in the window) (37). ALFF and fALFF values were extracted from CONN.

Metric	Description	Interpretation in rs-fMRI
Mean	The average of the BOLD signal over the course of the timeseries	The average % change in BOLD response over the time series.
Standard Deviation	The dispersion of the BOLD signal around the mean	Representation of the variation of % change in BOLD response from the mean throughout the time series.
Sample Entropy (SampEn) (37)	Generally defined as the lack of order/predictability in a system.	A measure of disorder across the BOLD time series, i.e. is the signal predictable or is it “random”. Can also describe the complexity of the signal.
Lyapunov Exponent (LyaExp) (38)	Measures the rate of divergence of trajectories in phase space. Can be used a measure of chaos in the system.	Describes how chaotic the BOLD signal is, i.e. how dependant future time points are on the present response.
Amplitude of Low-Frequency Fluctuations (ALFF) (39)	The amount of low-frequency components (0.001-0.01Hz) in a temporal signal’s frequency spectra.	Low-frequency fluctuations of the BOLD response are thought to correlate with spontaneous neuronal activities.
Fractional ALFF (fALFF) (39)	A normalized version of ALFF, where the amount low-frequency components are divided by the total number of frequency components.	Similar to ALFF, however due to the normalization fALFF is thought to provide a more robust measure for group analysis, however at the expense of sensitivity at the individual level.

Data analysis

Random forest models were implemented in Python (version 3) using the Scikit-Learn (40) and Imbalanced-Learn (41) libraries. To prepare the data for random forest analyses, the dataset was split into an 80/20 ratio, reserving 20% of the data for a final test. The remaining 80% was then further split into a 75/25 ratio, reserving 25% for a validation set. This creates an overall split of the data as 60/20/20. Of note is the data imbalance. All divisions were made preserving the original ratio of concussion to controls as much as possible, by performing random splits within each pool of data (i.e., concussion and healthy).

Models were trained using the 60% (n = 209) training set and evaluated on the 20% validation set (n = 81) to search for optimal parameters. The optimal parameters changed slightly depending on the metric being evaluated. The primary model parameter to be tuned was the number of estimators. The weighting was set to a 15:1 ratio to mirror the approximate ratio of the data imbalance. The decision trees were trained on the basis of maximizing the Gini Gain (with “gini” selected as the criterion upon which split quality was assessed in `scikit-learn`). Once the optimal parameters were determined, a finalized model was trained on the combined training and validation sets.

Performance was assessed on the final test set (n = 81) and input features were ranked and weighted by importance. After the feature ranking, the bottom 20% weighted features were removed, and this process was iteratively repeated until less than 10 features remained. This was then repeated for all six metrics and one more time to train an overall classifier using features from all metrics.

Our results are summarized in three ways. First, we present confusion matrices and ROC curves to report the accuracy of each random forest classifier with respect to discernability between concussion and healthy rsfMRI data. Second, we present a table to report the various anatomical regions that were shown to be most discriminatory between groups (i.e., concussion vs. healthy) for each metric included in our analysis. Third, we depict the findings reported in the aforementioned table on a standard space structural MRI, to visually display which anatomical regions (across all metrics studied) were most discriminatory between concussion and healthy subjects.

Model performance

Random forest models were run for each rsfMRI metric and are summarized in Figure 1 (with Panel A displaying the confusion matrix and ROC curve for both the mean and standard deviation metrics, given they performed similarly). Overall, area under the curve (AUC) ROC values ranged from 0.869 (for fALFF) to 0.993 (for ALFF), with most models peaking in accuracy after the third iteration (i.e., after removing the bottom 20% of features after both the first and second iterations and having run the model for the third time). We also ran a model (Panel F) with all metrics included, which reached an ROC AUC of 0.996. The purpose of building a model with all six metrics was to have all metric-ROI pairs (i.e., 6 metrics x 164 ROIs) in one classification model, to allow us to evaluate which metric-ROI pair (among all available) would be most discriminatory. From our data we suggest that of the six metrics analyzed, ALFF is the most discriminatory individual metric between concussion and healthy rsfMRI data.

Discriminatory regions

For each metric (and for the model incorporating all metrics combined), we aimed to identify the 10 most important ROIs (from the most accurate iteration, based on overall model accuracy on the test set) to model classification (Table 2). More specifically, for the iteration that yielded the most accurate model, we then identified the ROIs corresponding to that iteration that were most important in aiding the model discriminate between healthy and concussed rsfMRI data.

Metric	ROI	Importance / Gini index
Mean/Standard deviation	Frontal Medial Cortex	0.099
	Subcallosal Cortex	0.090
	Cerebellum 4 5 Left	0.077
	Planum Temporale Left	0.065
	Inferior Temporal Gyrus, posterior division Left	0.056
	Vermis 4 5	0.051
	Vermis 1 2	0.051
	Cerebellum 4 5 Right	0.050
	Cerebellum 10 Left	0.050
Sample entropy	Subcallosal Cortex	0.125
	Cerebellum 4 5 Left	0.097
	Cerebellum 4 5 Right	0.096
	Planum Temporale Right	0.077
	Planum Temporale Left	0.069
	Temporal Fusiform Cortex, posterior division Left	0.058
	Amygdala Right	0.057
	Inferior Temporal Gyrus, anterior division Left	0.039
	Pallidum Left	0.038
	Occipital Fusiform Gyrus Left	0.035
Lyapunov exponent	Hippocampus Right	0.171
	Hippocampus Left	0.118
	Subcallosal Cortex	0.115
	Parietal Operculum Cortex Left	0.103
	Parahippocampal Gyrus, posterior division Left	0.097
	Superior Temporal Gyrus, posterior division Right	0.087
	Temporal Occipital Fusiform Cortex Right	0.082
	Middle Frontal Gyrus Left	0.079
	Cerebellum Crus 2 Left	0.075
	Putamen Left	0.074
ALFF	Occipital Fusiform Gyrus Left	0.177
	Superior Frontal Gyrus Right	0.085
	Putamen Right	0.081
	Putamen Left	0.074
	Frontal Pole Right	0.067
	Frontal Operculum Cortex Right	0.062
	Vermis 8	0.050
	Vermis 7	0.047
	Central Opercular Cortex Left	0.044
	Heschl's Gyrus Right	0.043
fALFF	Lingual Gyrus Right	0.164
	Frontal Operculum Cortex Left	0.163
	Precentral Gyrus Right	0.141
	Thalamus Right	0.118
	Inferior Frontal Gyrus, pars opercularis Right	0.081
	Lingual Gyrus Left	0.081
	Supracalcarine Cortex Right	0.081
	Frontal Pole Right	0.073
	Cerebellum 7b Left	0.058
	Cerebellum 6 Left	0.039
All metrics	Putamen Right (ALFF)	0.104
	Occipital Fusiform Gyrus Left (ALFF)	0.094
	Putamen Left (ALFF)	0.070
	Frontal Pole Right (ALFF)	0.056
	Supracalcarine Cortex Left (ALFF)	0.053
	Vermis 7 (ALFF)	0.051
	Lingual Gyrus Right (fALFF)	0.044
	Precentral Gyrus Right (ALFF)	0.043
	Pallidum Right (ALFF)	0.035
	Cingulate Gyrus, anterior division (ALFF)	0.035

Table 2: Tabular summary of the ten most important anatomical regions that contributed to classification based on a given rsfMRI metric. For the “All metrics” sub-section, the ROI is first listed, followed by the corresponding rsfMRI metric in brackets.

Some ROIs were discriminatory between concussion and healthy subjects across multiple metrics. This would suggest that the BOLD signal in select ROIs can be discriminatory on a group-wise basis, irrespective of the metric computed. These results are presented visually in Figure 2. In sum, the three regions which were discriminatory across multiple metrics included the cerebellum, vermis, and putamen left, presenting with high important in 5, 4, and 4, models, respectively. It should be noted that sub-regions of the cerebellum and vermis are reported on in Table 2. However, for visual illustrative purposes, all sub-regions that comprised either the cerebellum or vermis were collapsed (e.g., with respect to ALFF, the sub-regions of vermis 7 and vermis 8 were simply categorized as vermis for the purposes of plotting Figure 2).

Our study took aim to understand which rsfMRI metrics (and anatomical ROIs) are most discriminatory between children with concussion and healthy controls. With respect to individual metrics, we found that ALFF was the most discriminatory, with an ROC of 0.993. As well, in the model that included all metrics (Fig. 1, panel F and the last sub-section of Table 2), 9 of the top 10 most important ROIs were paired to ALFF, altogether accounting for over 55% of the “importance”. This model with all metric-ROI pairs would also support the notion that ALFF is discriminatory between concussion subjects and healthy controls. In terms of ROIs, we found that the BOLD signal from the cerebellum, vermis, and putamen (left) discriminated between groups in 5, 4, and 4, models respectively. This suggests that these ROIs are discriminatory between groups, across multiple metrics. To our knowledge, this is the first paper to use random forest models to identify rsfMRI metrics and ROIs that are most discriminatory, at a group-level, in concussion.

The role of rsfMRI in concussion diagnostics and monitoring is increasing within the general population and a sports context. The BOLD signal is complex, and it can be impacted by concussion in different ways (42). For example, concussion can impact cerebral blood flow (CBF) which is a contributing factor to the BOLD signal (43, 44), and recovery of CBF is regionally variable, with some regions, such as the dorsal midinsular cortex, recovering slower in athletes who experienced delayed recoveries (44). It has also been recognized that concussion can impact synchronous brain activity (42), and metric selection should be sensitive to such neurophysiological changes. The post-injury changes to brain activity synchronization may help explain why our study identified ALFF, which is a measure of spontaneity, as the most discriminatory rsfMRI metric between concussion and healthy subjects. Future studies should also measure CBF in addition to BOLD metrics to understand their association, and whether certain rsfMRI metrics better track changes in brain blood flow after a concussion. Relating such changes to clinically accessible outcomes, such as symptoms or perhaps measures of autonomic dysfunction such as heart rate variability, should also be considered.

Further, with respect to the BOLD signal, it is important to note that both temporal sampling and total acquired volumes will directly impact which metrics can be extracted from the signal, and the quality of those metrics. Alongside the metrics chosen in this study, the BOLD signal can be analyzed in other ways, such as Hurst exponent (22) fractal dimension (23), or wavelet decomposition (45). However, signal fractality and Hurst exponent generally require longer acquisitions to be more accurate. Given our data included 180 timepoints, other metrics (such as the Lyapunov exponent and sample entropy) were used to provide a measure of complexity. The metrics which we selected for this study represented a variety of methods in which to interpret the BOLD signal, and therefore allowed for assessment of which type of BOLD-based metric is most discriminatory between concussion and healthy data. We did, however, focus on measures of brain spontaneity, given that such metrics have been studied in neurological populations and changes to them can be readily captured in the resting state (37, 46). Brain entropy has also been associated with cognition and behaviour (47, 48), suggesting its clinical relevance. In particular, measures of spontaneous functional brain activity (such as ALFF and fALFF) have been used in concussion to detect groupwise differences (49, 50). These metrics were included as they demonstrate promise, in neurological populations at large, since it was initially proposed as metrics that are sensitive to abnormal brain activity (and fluctuations of brain activity) (39).

Recent systematic reviews have identified that the resolution of concussion symptoms does not necessarily indicate resolution of injury-related neuropathologies (4). As such, it is important for studies to include outcome measures beyond symptoms. rsfMRI is one such outcome measure that may be able to provide insight into neuropathological features of concussion, and recent reviews demonstrate its potential as a biomarker for mild brain injury (17, 51). The studies included in said reviews use various features of the resting state BOLD signal to differentiate between brain injury patients and healthy control in a group-wise manner. Our study adds to this evidence-base. First, we demonstrated that ALFF may be able to best discriminate between concussion subjects and healthy controls. These data would indicate that of the various BOLD signal-based rsfMRI metrics that are now available, the ALFF may be most appropriate for concussion (in that our random forest models showed that this held the most between-group discriminatory ability). Recent reviews in the field have thoroughly consolidated the evidence base and reported on which networks demonstrate differences in brain injury (17), but they have not examined how the BOLD signal itself should be analyzed to best assess groupwise differences. Other concussion studies have examined different aspects of the BOLD signal (such as variability in the BOLD signal (52) or its regional homogeneity (25), or even scale-free dynamics of functional brain activity (53)), but they have not compared one resting state BOLD signal metric to another. In doing so, our study can help identify which rsfMRI metrics are most discernable and applicable in concussion. Future studies should examine the associations between ALFF and symptoms, and understand (through longitudinal designs) whether symptom recovery maps on to normalization of ALFF.

The ROIs which were most discriminatory (i.e., the regions which were most commonly identified as being discriminatory across the six rsfMRI metrics studied) included the cerebellum, vermis, and putamen (left). The specific sub-regions of the cerebellum and vermis that had the greatest importance to the model at its most accurate iteration are reported in Table 2. Living meta-analyses from tools such as Neurosynth.org (from the National Institutes of Health), demonstrate that searching the term “brain injury” reveals that there is often impairment in the cerebellum, which is consistent with our findings. Our data can also help inform studies that require a priori regional selection, such as spectroscopic analyses. By focusing on regions which we have shown to be discriminatory between groups, concussion-spectroscopy researchers have precedent for voxel selection, which may help yield more insightful results.

We acknowledge a number of limitations to our work. First, the control data were not prospectively collected by the study authors. The lack of control over MRI acquisition is a limitation with respect to the control data. However, as stated in our methodology section, we did perform quality control checks on each control rsfMRI to ensure that it was appropriate for inclusion. Second, our groups are considerably imbalanced, which may influence the random forest classification model. Although we were not able to collect or acquire additional concussion data, we used an analysis approach (based on the Python package imbalanced learn, or `imblearn`, part of the `sklearn` toolkit) to address the under sampling in our data. Third, we did not examine all rsfMRI metrics that are now available and used in clinical research. There are many, and more that are continually being developed. The selection of metrics for our study was based on those we believe are commonly used and therefore suitable for comparison. Fourth, just as there are many metrics to describe rsfMRI, there are numerous machine learning models to build classifiers (54, 55). For example, neural networks and support vector machines (SVM) are popular choices that show great performance in a number of different related fields and applications. However, other models, particularly neural networks, are typically referred to as black-box models. It is difficult to discern how decisions are being made or to identify specifically which features are more important in the decision-making process. Therefore, random forests were selected for their transparency and ease of creating hierarchical rankings of feature importance. Follow up study would ideally confirm these findings with other models.

Our work helps address challenges faced by concussion-MRI researchers. Selection of a resting state BOLD metric is often exploratory or not informed by guidelines or prior evidence. Likewise, regions to focus on (in seed-to-voxel or ROI-to-ROI analyses) is often too exploratory. Our study provides precedent for focusing on specific metrics such as ALFF, as well as specific regions such as the cerebellum, vermis, and putamen (left). Future research should take aim at examining the discriminatory ability of ALFF in their personalized concussion patients. These findings and/or our analysis approach can also be applied to different concussion sub-groups (including older participants, those with multiple injuries, those with persistent symptoms) to understand the utility of our approach outside of pediatric concussion. We believe our findings provide important insight into how the resting state BOLD signal can best be used to understand differences between the healthy and concussion brain, and how regions such as the cerebellum, vermis, and putamen (left) should be further studied given their groupwise discriminatory potential. As the field moves towards a personalized medicine approach, our findings may be able to help concussion-fMRI scientists focus their research on aspects of the rsfMRI signal and select ROIs that may be most reflective of the neuropathology of concussion.

ABIDE-II Autism Brain Imaging Data Exchange II

ALFF Amplitude of low frequency fluctuations

BOLD Blood-oxygen-level-dependent

CBF Cerebral blood flow

fALFF Fractional amplitude of low frequency fluctuations

fMRI Functional magnetic resonance imaging

MRI Magnetic resonance imaging

ROI Region of interest

rsfMRI Resting state functional magnetic resonance imaging

Ethics approval and consent to participate

This study was approved by the Hamilton Integrated Research Ethics Board prior to any data collection. Assent was acquired from participants aged less than 16 years, with those aged older providing informed consent. Parents of participants also provided parental consent.

Consent for publication

Not applicable.

Availability of data and materials

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Competing interests

Dr. Noseworthy is the CEO and co-founder of the data analytics company TBIFinder, Inc., which focuses on personalized diagnostics of brain injury. The company did not contribute any form of support or data for this project. The other authors declare that they have no competing interests.

Funding

Operational funding for this study was secured from the Canadian Institutes of Health Research (CIHR; MOP 133527). The first author (B Sharma) received doctoral funding support from CIHR as well (CIHR-CGS-D #157864) and the Research Institute of St. Joe’s Postdoctoral Award.

Authors’ contributions

The study was conceived by BS and CN, who also pre-processed the data and extracted rsfMRI metrics. CZ developed and ran the machine learning models. MDN maintained oversight over all prospective concussion imaging. BS prepared the first version of this manuscript. All authors (BS, CN, CZ, and MDN) reviewed, edited, and approved the submitted version of this manuscript.

Acknowledgements

Not applicable.

Amoo-Achampong K, Rosas S, Schmoke N, Accilien YD, Nwachukwu BU, McCormick F. Trends in sports-related concussion diagnoses in the USA: a population-based analysis using a private-payor database. Physician Sportsmed. 2017;45(3):239–44.
Patricios JS, Schneider KJ, Dvorak J, Ahmed OH, Blauwet C, Cantu RC et al. Consensus statement on concussion in sport: the 6th International Conference on Concussion in Sport–Amsterdam, October 2022. British Journal of Sports Medicine. 2023;57(11):695–711.
Varner C, Thompson C, de Wit K, Borgundvaag B, Houston R, McLeod S. Predictors of persistent concussion symptoms in adults with acute mild traumatic brain injury presenting to the emergency department. Can J Emerg Med. 2021;23:365–73.
Kamins J, Bigler E, Covassin T, Henry L, Kemp S, Leddy JJ et al. What is the physiological time to recovery after concussion? Systematic review. Br J Sports Med. 2017;bjsports-2016-097464.
Esopenko C, Sollmann N, Bonke EM, Wiegand TL, Heinen F, de Souza NL, et al. Current and emerging techniques in neuroimaging of sport-related concussion. J Clin Neurophysiol. 2023;40(5):398–407.
Churchill NW. Neuroimaging of Concussion. Tackling the Concussion Epidemic: A Bench to Bedside Approach. Springer; 2022. pp. 43–70.
Rausa VC, Shapiro J, Seal ML, Davis GA, Anderson V, Babl FE, et al. Neuroimaging in paediatric mild traumatic brain injury: a systematic review. Neuroscience & Biobehavioral Reviews; 2020.
Narayana S, Charles C, Collins K, Tsao JW, Stanfill AG, Baughman B. Neuroimaging and Neuropsychological Studies in Sports-Related Concussions in Adolescents: Current State and Future Directions. Front Neurol. 2019/06/11 ed. 2019;10:538.
Schmidt C. Trends in major sports nutition markets and demographics: Understanding the consumer market [Internet]. Vol. 2017. Euromonitor international; 2014. http://www.benefiq.ca/eng/wp-content/uploads/2014/10/3-Chris-Schmidt.pdf.
Chamard E, Lichtenstein JD. A systematic review of neuroimaging findings in children and adolescents with sports-related concussion. Brain Inj. 2018;1–16.
McCrea M, Meier T, Huber D, Ptito A, Bigler E, Debert CT et al. Role of advanced neuroimaging, fluid biomarkers and genetic testing in the assessment of sport-related concussion: a systematic review. Br J Sports Med. 2017;bjsports-2016-097447.
Monk SH, Legarreta AD, Kirby P, Brett BL, Yengo-Kahn AM, Bhatia A, et al. Imaging findings after acute sport-related concussion in American football players: A systematic review. J Clin Neurosci. 2019;61:28–35.
Gumus M, Santos A, Tartaglia MC. Diffusion and functional MRI findings and their relationship to behaviour in postconcussion syndrome: a scoping review. J Neurol Neurosurg Psychiatry. 2021;92(12):1259–70.
Biagianti B, Stocchetti N, Brambilla P, Van Vleet T. Brain dysfunction underlying prolonged post-concussive syndrome: A systematic review. J Affect Disord. 2020;262:71–6.
Hall CN, Howarth C, Kurth-Nelson Z, Mishra A, Interpreting. BOLD: towards a dialogue between cognitive and cellular neuroscience. Vol. 371, Philosophical Transactions of the Royal Society B: Biological Sciences. The Royal Society; 2016.
Gauthier CJ, Fan AP. BOLD signal physiology: models and applications. NeuroImage. 2019;187:116–27.
Morelli N, Johnson NF, Kaiser K, Andreatta RD, Heebner NR, Hoch MC. Resting state functional connectivity responses post-mild traumatic brain injury: a systematic review. Brain Injury. 2021;35(11):1326–37.
Suárez LE, Markello RD, Betzel RF, Misic B. Linking structure and function in macroscale brain networks. Trends Cogn Sci. 2020;24(4):302–15.
Davis T, LaRocque KF, Mumford JA, Norman KA, Wagner AD, Poldrack RA. What do differences between multi-voxel and univariate analysis mean? How subject-, voxel-, and trial-level variance impact fMRI analysis. NeuroImage. 2014;97:271–83.
Lv H, Wang Z, Tong E, Williams LM, Zaharchuk G, Zeineh M, et al. Resting-state functional MRI: everything that nonexperts have always wanted to know. Am J Neuroradiol. 2018;39(8):1390–9.
Von Wegner F, Laufs H, Tagliazucchi E. Mutual information identifies spurious Hurst phenomena in resting state EEG and fMRI data. Phys Rev E. 2018;97(2):022415.
Dong J, Jing B, Ma X, Liu H, Mo X, Li H. Hurst exponent analysis of resting-state fMRI signal complexity across the adult lifespan. Front NeuroSci. 2018;12:34.
Dona O, Noseworthy MD, DeMatteo C, Connolly JF. Fractal analysis of brain blood oxygenation level dependent (BOLD) signals from children with mild traumatic brain injury (mTBI). PLoS ONE. 2017;12(1):e0169647.
Vedaei F, Newberg AB, Alizadeh M, Muller J, Shahrampour S, Middleton D, et al. Resting-state functional MRI metrics in patients with chronic mild traumatic brain injury and their association with clinical cognitive performance. Front Hum Neurosci. 2021;15:768485.
Meier TB, Giraldo-Chica M, Espana LY, Mayer AR, Harezlak J, Nencka AS, et al. Resting-state fMRI metrics in acute sport-related concussion and their association with clinical recovery: a study from the NCAA-DOD CARE Consortium. J Neurotrauma. 2020;37(1):152–62.
Danielli E. A multi-modal application of magnetic resonance imaging (MRI) techniques to identify and quantify brain abnormalities in retired professional football players. 2022.
Liang X, Wang J, Yan C, Shu N, Xu K, Gong G, et al. Effects of different correlation metrics and preprocessing factors on small-world brain functional networks: a resting-state functional MRI study. PLoS ONE. 2012;7(3):e32766.
Huotari N, Raitamaa L, Helakari H, Kananen J, Raatikainen V, Rasila A, et al. Sampling rate effects on resting state fMRI metrics. Front NeuroSci. 2019;13:279.
Andellini M, Cannatà V, Gazzellini S, Bernardi B, Napolitano A. Test-retest reliability of graph metrics of resting state MRI functional brain networks: A review. J Neurosci Methods. 2015;253:183–92.
Chu PP, Golestani AM, Kwinta JB, Khatamian YB, Chen JJ. Characterizing the modulation of resting-state fMRI metrics by baseline physiology. NeuroImage. 2018;173:72–87.
Sharma B, Nowikow C, DeMatteo C, Noseworthy MD, Timmons BW. Sex-specific differences in resting-state functional brain activity in pediatric concussion. Sci Rep. 2023;13(1):3284.
Di Martino A, O’connor D, Chen B, Alaerts K, Anderson JS, Assaf M, et al. Enhancing studies of the connectome in autism using the autism brain imaging data exchange II. Sci data. 2017;4(1):1–15.
Davis AD, Noseworthy MD. Motion and distortion correction of skeletal muscle echo planar images. Magn Reson Imaging. 2016;34(6):832–8.
Nowikow C, Sharma B, Noseworthy MD. P174. The lack of B0 maps in rs-fMRI databases: A big problem or no big deal? Book of abstracts ESMRMB. Basel, Switzerland: Springer; 2023. pp. S212–3.
Richman JS, Moorman JR. Physiological time-series analysis using approximate entropy and sample entropy. Am J Physiol Heart Circ Physiol. 2000;278(6):H2039–49.
Wang Z, Li Y, Childress AR, Detre JA. Brain entropy mapping using fMRI. PLoS ONE. 2014;9(3):e89948.
Nezafati M, Temmar H, Keilholz SD. Functional MRI signal complexity analysis using sample entropy. Front NeuroSci. 2020;14:520547.
Keilholz S, Maltbie E, Zhang X, Yousefi B, Pan WJ, Xu N, et al. Relationship between basic properties of BOLD fluctuations and calculated metrics of complexity in the human connectome project. Front NeuroSci. 2020;14:550923.
Zou QH, Zhu CZ, Yang Y, Zuo XN, Long XY, Cao QJ, et al. An improved approach to detection of amplitude of low-frequency fluctuation (ALFF) for resting-state fMRI: fractional ALFF. J Neurosci Methods. 2008;172(1):137–41.
Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al. Scikit-learn: Machine learning in Python. J Mach Learn Res. 2011;12:2825–30.
LemaÃŽtre G, Nogueira F, Aridas CK. Imbalanced-learn: A python toolbox to tackle the curse of imbalanced datasets in machine learning. J Mach Learn Res. 2017;18(17):1–5.
Sicard V, Harrison AT, Moore RD. Psycho-affective health, cognition, and neurophysiological functioning following sports-related concussion in symptomatic and asymptomatic athletes, and control athletes. Sci Rep. 2021;11(1):1–10.
Churchill NW, Hutchison MG, Graham SJ, Schweizer TA. Symptom correlates of cerebral blood flow following acute concussion. Neuroimage: Clin. 2017;16:234–9.
Meier TB, Bellgowan PS, Singh R, Kuplicki R, Polanski DW, Mayer AR. Recovery of cerebral blood flow following sports-related concussion. JAMA Neurol. 2015;72(5):530–8.
Patel AX, Kundu P, Rubinov M, Jones PS, Vértes PE, Ersche KD, et al. A wavelet method for modeling and despiking motion artifacts from resting-state fMRI time series. NeuroImage. 2014;95:287–304.
Xin X, Yu J, Gao X. The brain entropy dynamics in resting state. Front NeuroSci. 2024;18:1352409.
Wang Z. The neurocognitive correlates of brain entropy estimated by resting state fMRI. NeuroImage. 2021;232:117893.
Zhang S, Rogers BP, Morgan VL, Chang C. Association between fMRI brain entropy features and behavioral measures. Medical Imaging 2020: Image Processing. SPIE; 2020. pp. 185–94.
Zhan J, Gao L, Zhou F, Bai L, Kuang H, He L, et al. Amplitude of low-frequency fluctuations in multiple-frequency bands in acute mild traumatic brain injury. Front Hum Neurosci. 2016;10:27.
Shi J, Du X. Multi-modal analysis of resting-state fMRI data in mTBI patients and association with neuropsychological outcomes. Front Neurol. 2021;12:639760.
Puig J, Ellis MJ, Kornelsen J, Figley TD, Figley CR, Daunis-i-Estadella P et al. Magnetic resonance imaging biomarkers of brain connectivity in predicting outcome after mild traumatic brain injury: a systematic review. J Neurotrauma. 2020.
Anderson ED, Talukdar T, Goodwin G, Di Pietro V, Yakoub KM, Zwilling CE, et al. Assessing blood oxygen level–dependent signal variability as a biomarker of brain injury in sport-related concussion. Brain Commun. 2023;5(4):fcad215.
Churchill NW, Hutchison MG, Graham SJ, Schweizer TA. Scale-free functional brain dynamics during recovery from sport‐related concussion. Hum Brain Mapp. 2020;41(10):2567–82.
Khosla M, Jamison K, Ngo GH, Kuceyeski A, Sabuncu MR. Machine learning in resting-state fMRI analysis. Magn Reson Imaging. 2019;64:101–21.
Chen Q, Bi Y, Zhao X, Lai Y, Yan W, Xie L, et al. Regional amplitude abnormities in the major depressive disorder: A resting-state fMRI study and support vector machine analysis. J Affect Disord. 2022;308:1–9.

Competing interest reported. Dr. Noseworthy is the CEO and co-founder of the data analytics company TBIFinder, Inc., which focuses on personalized diagnostics of brain injury. The company did not contribute any form of support or data for this project. The other authors declare that they have no competing interests.

Download PDF

Editorial decision: Revision requested
14 Jun, 2024
Editor assigned by journal
14 Jun, 2024
Submission checks completed at journal
13 Jun, 2024
First submitted to journal
13 Jun, 2024

You are reading this latest preprint version

Which voxel-wise resting state fMRI metric is the most discriminatory for concussion? A secondary analysis.

Status:

Version 1

Abstract

Figures

INTRODUCITON

METHODOLOGY

RESULTS

DISCUSSION

CONCLUSION

Abbreviations

Declarations

References

Additional Declarations

Status:

Version 1