A Retrospective Study of Predictors of Return to Duty versus Medical Retirement in an Active Duty Military Population with Blast-Related Mild Traumatic Brain Injury

Subjects

We obtained a sample of retrospective data from 313 previously deployed service members with blast injury, seen at the ISCRC from 2009 through 2014. Neuropsychological reports were available for 61 of the subjects referred to MEB, which included an assessment of the validity of the cognitive tests. Ten subjects were excluded because it was concluded that the cognitive test results were invalid. This reduced the total sample size to N = 303.

Patients treated at the ISCRC are referred typically by their medical officer or other healthcare provider. In some cases, patients self-refer for evaluation. The center currently receives about 40 new referrals per month. A nurse practitioner or physician assistant as well as a team of nurses evaluate the patient initially and review the history of brain injury. Subsequently, patients are evaluated by an interdisciplinary team: speech and language, physical therapy, occupational therapy, behavioral health, neurology, sports medicine, and pastoral counseling. Providers from each discipline determine whether the patient requires their specific services based on their symptoms and functional impairment.

After the initial evaluations, the patient and providers meet as a team to discuss the recommended treatment plan and next steps. Evaluation and care are provided in a collaborative environment that promotes physical, psychological, and spiritual health. Interdisciplinary care is facilitated by the colocation of these providers in the ISCRC. Follow-up treatment team meetings are scheduled every four weeks until discharge to reassess progress in the program, make necessary adjustments, and plan for eventual discharge. The typical patient moves through the program over a 12-week period. Patients are offered traditional medical treatment as well as complementary medical programs including yoga, iRest^® meditation, cognitive behavioral therapy groups, acupuncture, mindfulness-based stress reduction, art therapy, therapeutic writing, equine facilitated communication, and therapeutic photography. According to internal ISCRC tracking data, approximately 80% of the service members who participate in the program return to full duty.

Information from clinical records, neuropsychological tests, and self-report measures were tested individually for association with MEB status. In addition, we derived quantitative measures of regional white matter integrity from the diffusion magnetic resonance imaging (MRI) images. Patient information was collected from a variety of sources. Some data were retrieved from the electronic Armed Forces Health Longitudinal Technology Application (AHLTA) medical record. The MRI data were retrieved from the picture archiving and communication system (PACS) at Camp Lejeune. Other information was retrieved manually from the patient's clinic notes, including reported symptoms and history of brain injury. All patient data were de-identified and coded with a study-specific identifier at Camp Lejeune before transfer to the University of Pennsylvania for analysis.

All of the variables in the data set have some degree of missing data. Patients seen more recently have more data because the use of standardized tests and questionnaires has expanded at ISCRC over the time span of the study. This means the statistical tests have different power because the sample size varies substantially. We identify the sample size for each test with a lower case n.

Population characteristics and medical histories

Limited demographic information and individual history were recorded where available. Patient demographics are summarized in Table 1. To protect patient privacy, demographic information was limited to sex, race, and marital status. Individual age was not recorded in the data set; based on internal clinic records, the average age of the population seen is 27.9 years with 48% of patients less than age 26. Before injury, all subjects were on full active duty and required to meet the physical and cognitive standards of active duty Marines.

Information from clinical histories

The patient history contains, in narrative form, self-reported individual and family medical history including history of TBI, reports of symptoms, and other information related to the patient recorded by the examining physician. Blast TBI history was determined by patient report in the medical record. Following standard definitions, a TBI was classified as mild if the reported loss of consciousness and/or confusion and disorientation was shorter than 30 min. A data structure was defined to encode information as either numerical (e.g., number of reported blast injuries) or categorical (e.g., was there a blast injury caused by an improvised explosive device [IED], Yes/No). Records were reviewed manually, and the extracted data was recorded in a spreadsheet.

We recorded the number of head traumas from blast and lifetime total concussions, including injuries sustained before military service. In addition, we recorded details about blast exposures when they were present in the clinical notes, including devices of blast injury, the estimated proximity to the device, and the location of the individual at the time of injury.

Standardized tests and surveys

In addition to the narrative history, standardized tests and surveys were administered to some patients. Most of these have now become standard at the clinic. The tests included in the study are:

MRI data

Neuroimaging data was acquired on a 1.5T Siemens Espree scanner. The standard neuroimaging protocol at the TBI clinic includes sagittal T1 and axial T2 and fluid attenuation inversion recovery (FLAIR) images and an axial diffusion-weighted imaging series. Diffusion was measured along three orthogonal planes, at b-values of (0, 500, and 1000) s mm². The earliest FLAIR data (2009–2010) had a resolution of 0.94 × 0.94 × 6.5 mm³ (n = 62). The majority of images after 2010 had an in-plane resolution of 0.45 × 0.45 mm² (n = 167), but some were acquired at intermediate resolutions of approximately 0.75 × 0.75 mm² in plane and 6.5 or 5.2 mm-thick slices (n = 43). The diffusion images were almost all acquired at similar resolution, either 1.25 × 1.25 × 6.5 mm³ (n = 137) or 1.2 × 1.2 × 6.5 mm³ (n = 113), with a few later scans acquired at the highest resolution of 0.6 × 0.6 × 5.2 mm² (n = 22). In all cases, apparent diffusion coefficients (ADC) were computed on the scanner, and these were used for analysis.

To analyze the ADC in white matter, we transferred regions of interest (ROIs) from a standard template brain to each subject brain by image registration, using the open-source Advanced Normalization Tools (ANTs) toolkit. ²⁸ The labels are defined in the International Consortium of Brain Mapping-diffusion tensor imaging-81 (ICBM-DTI-81) atlas and identify major white matter tracts in the MNI152 brain template space. ²⁹ Four ROIs were selected to test for association with MEB status: the genu of corpus callosum, the splenium of corpus callosum, the posterior limb of the internal capsule (bilateral), and the anterior corona radiata (ACR) (bilateral). These regions have been found to have anisotropy changes in multiple DTI studies of TBI. ³⁰ Other white matter ROIs including the cingulum and uncinate fasciculus were also abnormal in several studies in the meta-analysis, but the size and curvature of these structures made them too difficult to reliably segment in our data.

Transferring labels from the MNI152 template to individual subjects involves image registration of images with very different contrasts and spatial resolution. To meet this challenge, we broke the registration problem down into a series of simpler registrations, which were combined to map the white matter ROIs to each individual image. The first step was the construction of a population-specific template (Fig. 1). The population template was constructed by iteratively registering and averaging FLAIR images from 100 randomly chosen subjects. ³¹ The population template has 1 mm³ spatial resolution, like the MNI152 template: by averaging information from a large number of subjects, we were able to produce a reasonable interpolation between the 6.5-mm–thick slices of the individual images.

OPEN IN VIEWER

FIG. 1.

The registration pipeline for apparent diffusion coefficient (ADC) analysis. (a) The population template is built from 100 randomly selected fluid attenuation inversion recovery (FLAIR) images. (b) The registration pipeline computes a series of transforms that are combined to warp ADC and white matter regions of interest to the subject space.

Each individual structural image was then registered to the population template. Because the population template is an average of images from the same population of subjects, using the same MRI imaging sequence and scanner, a population template helps accomplish unbiased and robust registration. ³² The registration from the population template to MNI152 was computed once and applied consistently for all subject images.

The final registration shown in Figure 1 aligns the ADC image from each subject to the subject's FLAIR image. The two images are already closely aligned because they were acquired at the same time. A small correction accounts for patient motion and geometric distortions that differ between MRI sequences.

The subject FLAIR images were segmented into three tissue classes: cerebrospinal fluid (CSF), gray matter, and white matter, using the Atropos tool in ANTs. ³³ Each slice of the FLAIR image was segmented independently. A k-means classifier with three classes was used to initialize the segmentation. After each slice had been segmented, the mean intensities of gray matter, white matter, and CSF were computed over the whole brain, and the slice-by-slice segmentation was repeated with the k-means initialized consistently for each slice. On visual inspection, the segmentation appeared successful at segmenting CSF voxels from those containing gray or white matter; however, the separation of gray from white matter voxels was less reliable, especially in the internal capsule ROI (Fig. 2). Therefore, we did not exclude voxels within the white matter ROIs (defined by registration from the MNI template) if they were segmented as gray matter in the FLAIR image.

OPEN IN VIEWER

FIG. 2.

Region of interest (ROI) post-processing. The labels are first warped to the subject image from the template. The warped labels are refined by removing voxels where the apparent diffusion coefficient (ADC) is above a maximum threshold, or where the segmentation of the fluid attenuation inversion recovery (FLAIR) image classifies the voxel as cerebrospinal fluid.

After registration and segmentation of the native FLAIR images, we resampled the FLAIR images to a consistent resolution of 0.65 × 0.65 × 6.5 mm to minimize any bias in the resampling of the template ROIs into the subject FLAIR images. The registration warps were composed and applied once, transferring the white matter ROIs from MNI152 to the resampled subject FLAIR image with a single interpolation. The distortion-corrected ADC images were also resampled into the same FLAIR image space.

The ROIs were then masked with the FLAIR segmentations to remove voxels classified as CSF, and further masked with an ADC threshold to remove voxels with partial CSF volume. The ADC in normal white matter is relatively stable around 700 μm² ms⁻¹ , and we removed voxels where the ADC was greater than 1000 μm² ms⁻¹ (Fig. 2). To verify that this threshold was well above normal white matter variance, we computed the group mean of each individual's median ADC in labeled voxels before masking (748 μm² ms⁻¹ ), and the mean interquartile range (IQR) (110 μm² ms⁻¹ ). The final masked labels were reviewed manually, and six of 278 ADC images were removed from the analysis because one or more of the ROI labels were of low quality because of registration or segmentation errors.

Statistical analysis

The main aim of this work was to construct a multivariate model that could predict MEB or RTD outcomes. Because currently there is no single prognostic indicator for mTBI, a variety of data is collected and assessed by the ISCRC medical team; however, these have not been combined previously into a multivariate statistical model. We hypothesized that a multivariate model could predict MEB or RTD outcomes. Models with too many variables, however, are liable to overfit to training data and generalize poorly. We therefore performed exploratory univariate tests and selected variables based on the univariate results as well as the availability and completeness of data and previous evidence in the TBI literature. The p values of the univariate tests are uncorrected for multiple comparisons.

Dichotomous categorical variables were tested for association with MEB status, using the “epiR” package (https://CRAN.R-project.org/package=epiR) in R (http://r-project.org). We report the incidence risk ratio (IRR, also called relative risk) of MEB referral between dichotomous groups, with a 95% confidence interval (CI) as well as a p value from a chi-square test for differences in outcome between groups with and without exposure to a hypothesized risk factor.

For example, consider a test of whether marital status is correlated with the incidence of MEB referral. Among married patients, 46 were referred to MEB and 77 RTD. The incidence risk of MEB is therefore 46 / (46+77) = 37.4%. The incidence risk differs from the odds (46/77 = 0.60), although the two become similar when the incidence risk is small. Among nonmarried patients, 21 were referred to MEB and 51 RTD, an incidence risk of 29.2%. This results in an IRR of 1.28, suggesting that married patients are more likely to be referred to MEB. The estimated 95% CI for the IRR is (0.84, 1.96), however, which spans 1.0, suggesting that the effect is not significant. The corresponding p value is 0.24, and we would not reject the null hypothesis that marital status is uncorrelated with MEB.

In the results section, we report the total numbers in the MEB and RTD group (nM and nR), the IRR, CI, and associated p value. For the example above, this would be nM = 67, nR = 128, IRR = 1.28, CI = (0.84, 1.96), p = 0.24. Full incidence risk data are listed in Table 2. Point estimates and CIs for both the IRR and odds ratio may be computed from this table using the “epi.2by2” function in epiR.

For continuous data, we performed two-sample, two-tailed t tests, without assuming equal variance between the MEB and RTD groups (Welch t test). The p values are reported directly by the “t test” function in R. The t statistics are positive where the sample mean of the MEB group μM is greater than that of the RTD group μR.

Finally, we combined variables of interest to predict MEB or RTD status using a binomial logistic regression, implemented in the “glm” method in R. In this model, the dependent variable is MEB or RTD status, encoded as 1 for MEB and 0 for RTD. The model estimates the probability of MEB referral given the variables of interest.

RTD and MEB referral outcomes

Of the 303 subjects, 85 were referred to MEB and 178 RTD. The remaining 40 did not have a definite status at the time of data collection, either because they were still in treatment or because they had left the military without being medically retired. Among the 85 subjects referred to MEB, the specific medical grounds for referral were recovered for 84 subjects, coded as TBI-neurological, psychiatric, orthopedic, or other reasons. The codes are summarized in Table 3. The TBI-neurological code was defined to include nonpsychiatric neurological diagnoses attributable to TBI. Psychiatric diagnoses, mainly PTSD, were coded separately. All 84 subjects had either a TBI or psychiatric diagnosis, and in 77 cases, one of these was the primary diagnosis.

Number of TBI

All service members included in this study have at least one blast-related TBI. The number of TBI from blast and nonblast mechanisms was gathered from patient self-reports as recorded in clinic notes and is listed in Table 1. Injury mechanisms other than blast were also relatively common—145 subjects reported at least one TBI not caused by a blast. The median total TBI was three, with IQR one to four.

Subjects with three or more total TBI, including blast and nonblast injuries (3+ TBI) were more likely to be referred to the MEB than those with a single TBI: nM = 66, nR = 140, IRR = 1.64, CI = (1.0, 2.69), p = 0.039. This particular test was based on an earlier study of active duty US Marines by Spira and colleagues ³⁴ who found that three or more lifetime concussions was associated with worse post-concussive symptom reporting compared with one or zero concussions. We also compared subjects with 3+ TBI with those with one or two TBI, which was not significant: nM = 85, nR = 178, IRR = 1.34, CI = (0.93, 1.92), p = 0.11.

Characteristics of blast injury

Many of the clinic notes included limited information about the circumstances of blast-related injuries, which we hypothesized may be related to outcomes. We recorded information on blast device, estimated distance from the device, and the location of the service member inside or outside a vehicle when the blast occurred (Table 2). Of 290 subjects with information about the blast device, the most common was an IED (254 subjects), followed by rocket propelled grenades (39 subjects). Because of the predominance of IED exposure in the data set, we did not perform statistical tests on blast device. Compared with the largest group, “vehicle enclosed,” dismounted or vehicle open location was not related to MEB (p = 0.53 and p = 0.99, respectively). There was no association with closest proximity to a blast device and MEB status (p = 0.92). The actual severity of the blast injury depends on many factors not accounted for in these tests, including the explosive power of the device. Most service members reported injuries from IEDs, a category that encompasses a wide range of explosive types and sizes.

Standardized questionnaires

Of the standardized test scores, only RBANS was significantly associated with MEB status. The results of the tests are summarized in Table 4. In addition to potential utility in predicting MEB referral, these tests offer some insight into the physical and psychological symptoms of clinic patients seeking treatment at the ISCRC. We briefly discuss each test in turn, with reference to diagnostic scales where available. The sample size and group averages below are for all patients in the data set, including those without a defined MEB or RTD status at the time of data collection. The tests in Table 4 include only those subjects with a defined MEB or RTD status.

RBANS n = 117

the mean score was 84.6. The RBANS scoring algorithm includes adjustment for subject age and is calibrated such that the mean result is 100 in a US adult civilian population.

The scores were significantly lower in the MEB compared with the RTD group: (nM = 29, nR = 56, μM = 79.14, μR = 88.45, t = −3.35, p = 0.0013). One subject in the RTD group had an unusually high score (RBANS = 138, z = 3.71); however, the effect was robust to the removal of this point (p = 0.002). We hypothesized that the difference between the groups might be concentrated in particular cognitive domains within the RBANS. To test this, the individual cognitive domain scores were recovered where available, n = 75 subjects. We ran t tests on each domain score; the results are listed in Table 4 and plotted in Figure 3. Attention and delayed memory domain scores were significantly different between the MEB and RTD groups.

OPEN IN VIEWER

FIG. 3.

Repeatable Battery for Neuropsychological Status (RBANS) overall performance and domain scores. RBANS total scores were significantly lower for the medical evaluation board (MEB) group. Within individual domains, attention and immediate memory were significantly lower in the MEB group. Boxes show interquartile range (IQR) with a black line indicating the median. Whiskers extend to 1.5 * IQR; points outside this range are shown as black circles.

Multivariate model of MEB or RTD outcome

The predictor variables for the logistic regression were chosen based on the univariate tests: RBANS, ADC in the ACR, and a history of three or more TBI. To control for ADC image resolution, we regressed ADC against image resolution using all available ADC data and used the residuals as input to the logistic regression. The sample size for the logistic regression was n = 81. Although three or more TBI did not correlate significantly with MEB in the univariate analysis (except when subjects with two TBI were excluded), we included it in the logistic model because it trended in the expected direction (more head injuries associated with an increased likelihood of MEB), and previous studies have implicated multiple traumas as a risk factor for poor recovery from mTBI.

As in the univariate analysis, RBANS performance is the most significant predictor (β = −0.073, z = −2.96, p = 0.0031). Three or more head trauma was also a significant predictor (β = 1.20, z = 2.13, p = 0.033), but the ACR ADC was not significant (β = 0.013, z = 1.11, p = 0.27). To evaluate the logistic model as a classifier, we computed a receiver operating characteristic (ROC) curve with the “pROC” package ³⁶ in R. The ROC curve plots sensitivity against (1 – specificity) for different classifier thresholds of the logistic regression output, as shown in Figure 5.

OPEN IN VIEWER

FIG. 5.

Logistic regression output used as a classifier. The curve shows the specificity and sensitivity of different classification thresholds. The highlighted point has the maximum sum of sensitivity and specificity along the curve. ROC, receiver operating characteristic.

The highlighted point on the curve is placed at the point that maximizes the sum of sensitivity and specificity. At this threshold, the sensitivity is 0.815, meaning 81.5% of the subjects who went to MEB are classified as MEB. The specificity is 0.667, meaning 66.7% of subjects who RTD are correctly classified as RTD. The area under the curve (AUC) is 0.77. A perfect classifier has an AUC of 1.0, while a classifier performing no better than chance has an AUC of 0.5. A CI for the AUC was computed using bootstrap resampling of the classifier data; the 95% CI was (0.66, 0.86) after 2000 stratified bootstrap iterations.

We further tested the AUC by permutation testing. For each of 2000 trials, the MEB data were permuted, and the same logistic model was fit to the permuted data and used as a classifier. We define the p value for this test as the fraction of permuted data sets where the ROC AUC exceeded the value in the actual data. The permuted AUC exceeded the actual AUC in 4/2000 iterations, p = 0.002. To illustrate the relative importance of each variable, we computed AUC for models with one term left out (Table 6). The AUC is highest when the ADC is left out (AUC = 0.76) and lowest when RBANS is left out (AUC = 0.67).

Neuroimaging results

Previous studies have shown differences in white matter between mTBI and control populations, including mTBI caused by blast. ^38,39 In a study of military veterans with blast-related mTBI, Miller and colleagues ⁴⁰ found a significant correlation between distributed white matter abnormalities and post-concussive symptoms. In this study, we found significant differences related to RTD or MEB status only in the ACR, and the effect size was small (Table 5). The ADC images are quite limited in their ability to quantify the diffusion in white matter, compared with DTI. ⁴¹ Because of the low spatial resolution of the ADC images, we limited our analysis to four regions of interest. The choice of regions was guided by the TBI literature; however, these previous studies were mostly on civilian populations, and it is possible that the spatial distribution of white matter injuries is different in the military population exposed to blast. Diffusion anisotropy in the ACR has been linked to attentional control. ⁴² In the present study, the RBANS results showed significant correlation between MEB and attention (p = 0.0023). It is possible that the ACR ADC correlation with MEB is also related to attention; however, ACR ADC was not directly correlated with RBANS attention scores (p = 0.27).

Defining the relationships between white matter abnormalities, domain specific cognitive impairments, and a complex measure of functional recovery like readiness for military duty remains a challenging problem. Further exploration of these trends is planned for future studies, using improved diffusion imaging to better characterize and localize white matter pathology.

Multiple concussions

Spira and coworkers ³⁴ found that Marines with three or more lifetime concussions reported lower performance or worse symptoms on a range of cognitive and post-concussive tests, when compared with service members with one or zero concussions. The population in the study by Spira and coworkers ³⁴ was also active duty Marines but included Marines with no history of TBI, and excluded Marines not on full duty. We performed a similar analysis by comparing service members with one TBI with those with three or more, and we found that multiple mTBI is associated with a reduced likelihood of RTD. One limitation of the use of number of mTBI in prognosis is that we must rely usually on patient recall of the number of injuries and their severity. More complete reporting and the adoption of reliable wearable technology to measure mTBI exposure may improve prediction.

RBANS performance

In the univariate tests, RBANS score was the strongest predictor of MEB versus RTD outcome. The sample size of subjects with RBANS data was smaller (n = 85) than for tests of neuroimaging (n = 239) or 3+ vs. 1 TBI (n = 206). It remained the strongest result in the logistic model, however, including the (n = 81) subjects who had data for RBANS, 3+ TBI, and neuroimaging.

Post-concussive symptoms, depression, and PTSD

There is growing recognition in the literature that post-concussive and psychological symptoms frequently co-occur, presenting distinct challenges to both diagnosis and treatment. ^34,37,43,44 In this study, we had data for the HIT-6, NSI, the PCL-M, PHQ-9, and PSQI on a subset of the patients. Scores on these tests did not correlate with RTD status, but they show that many service members in both the RTD and MEB groups are affected by symptoms of headache, depression, sleep disorder, and post-traumatic stress.

Multivariate model

A logistic regression model was used to classify individuals as RTD or MEB according to RBANS, ADC, and a history of multiple (≥3) concussions with approximately 82% sensitivity and 67% specificity. The accuracy of this model was evaluated only via bootstrap and permutation testing in the same data set, and the sample for this model was relatively small, n = 81. An independent sample is required to test generalizability of the model. Variable selection based on univariate tests may lead to the inclusion of noninformative variables because of type I error, or the exclusion of variables that would be predictive in the context of a multivariate model. Missing data also complicated the selection of variables and precluded much exploration of different multivariate models. With a larger sample of complete multivariate data, machine learning techniques could be applied to train and test a more powerful predictive model.