Modeling Progressive Brain Volume Losses in Chronic Moderate Severe Traumatic Brain Injury: Extensive, but Attenuating

ABSTRACT

Objective

Progressive brain atrophy has now been widely observed in the chronic stages of moderate–severe traumatic brain injury (msTBI), but there is only coarse information regarding its prevalence and magnitude, and whether rate of atrophy accelerates or attenuates across time. We examined these questions in a Canadian inception cohort from the Toronto Rehab TBI Recovery Study.

Methods

N = 123 patients with post‐traumatic amnesia ranging from > 1 h to 10 weeks (70%, > 7 days) were administered up to three MRIs at approximately 5 months, 1 year, and 3 years post‐injury. N = 32 healthy controls were scanned up to two times on average 14 months apart. Ventricle‐to‐brain ratio (VBR) was calculated and manual segmentation of the hippocampi (HPC), corpus callosum (CC), and substructures was employed. For each region, annualized rates of progressive atrophy were modeled from approximately 5 months to 1 year and 1 to 3 years post‐TBI; proportions of patients with annualized volume losses of 2.5%–5%, 5%–10%, and > 10% were estimated.

Results

Mixed‐effects longitudinal models estimated robust progressive atrophy in both time windows in all regions compared to controls, but with significant attenuation across windows. In the earlier window, > 10% annualized loss was estimated in 76.2% of patients for VBR, 53.2% for CC, and 25.0% for HPC bilaterally. In the later window, at least 5%–10% annualized loss was estimated in 63.3% of patients for VBR, 49.4% for CC, and 17.7% for right HPC; for left HPC, 6.3% of patients showed at least 2.5%–5% annualized decline.

Interpretation

Clinically concerning prevalence and degree of progressive volume losses were observed in the chronic stages of msTBI, with tapering of losses across time, and regional variations. Research is needed to understand and mitigate clinical impact of these declines.

1. Introduction

Approximately 20% of traumatic brain injuries fall into the moderate–severe range (Glasgow Coma Scale score < 13; loss of consciousness of ≥ 30 min; post‐traumatic amnesia [PTA] ≥ 1 day) [1], which is associated with life‐altering cognitive, physical, and psychosocial dysfunction [2]. Although there is growing understanding that moderate–severe traumatic brain injury (msTBI) is a chronic disease process [2, 3, 4], there is much less discussion in the scientific literature that it can also be progressive. However, deteriorative changes in brain volume, white matter integrity, and functional connectivity commencing in the early months and years post‐injury have been observed in both younger and older adults by our group and others [5, 6, 7, 8, 9, 10, 11, 12, 13], as have progressive declines in cognitive function and increasing disability [14, 15, 16, 17, 18].

These deteriorative findings are concerning given that smaller brain volume is associated with reduced brain reserve (which buffers against aging‐related neuronal losses), and with earlier onset of dementia [19]. To study and manage these effects, an understanding is needed of the frequency and degree of these progressive declines [8] and whether the rate of decline is getting worse, better, or stabilizing over time.

We have begun to examine these questions in a Canadian inception cohort, The Toronto Rehabilitation Institute TBI Recovery Study [5, 6, 11, 20, 21], comprising patients with TBI ranging from moderate to very severe followed up from ~2 months to 3+ years post‐injury and recruited from an in‐patient neurorehabilitation program. We previously found that 96% of patients showed significantly greater progressive brain atrophy than healthy controls in at least one of the following regions: bilateral hippocampi (HPC), corpus callosum (CC), or whole brain (ventrical to brain ratio [VBR]) from approximately 5 to 20 months post‐injury [20] (measured by monthly percent change). In an overlapping sample of patients and focusing on subcortical regions (nucleus accumbens, amygdala, brainstem, HPC, thalamus), we examined decline from ~5 to 12 versus 12 to 30+ months post‐injury. Employing exploratory principal components analysis followed by an examination of nonlinear trends [6], we did not find evidence for attenuation (plateauing) of atrophy, but rather we found comparable slopes at both time windows.

Building on these studies, the first objective of the current study was to revisit the question of rate of decline using a more powerful statistical approach. The current approach allowed us to estimate different average rates of decline in different periods of time post‐injury, to adjust these rates for the expected age‐related decline exhibited in control subjects, and to estimate the net variability in the underlying rate of decline in the post‐injury population, above and beyond variability attributable to measurement error. The second objective was to estimate the proportion of patients at risk of progressive volume loss in these structures and the magnitude of that loss. To this end, we modeled the proportion of patients showing progressive volume loss during the earlier and the later time windows at three stratifications of deterioration: annualized volume loss of 2.5%–5%, 5%–10%, and > 10%. Both objectives focused on the whole brain (VBR), and manually segmented HPC (including the head, body, and tail) and CC (including the genu, body, and splenium).

2. Subjects and Methods

2.1. Subjects

2.1.1. TBI Patient Group

Informed consent was obtained for each study participant. For the current study, N = 123 patients with clinically confirmed TBI were drawn from the larger Toronto Rehabilitation Institute TBI Recovery Study database. Participants in the parent study were recruited from a large, urban, in‐patient brain injury program in Toronto; all patients had injuries severe enough to warrant in‐patient neurorehabilitation, and the majority of the sample were at least severe based on PTA. The prospective parent study entailed longitudinal MRI and neuropsychological assessments. Key inclusion/exclusion criteria for the parent study were (1) PTA of > 1 day and/or Glasgow Coma Scale score of 12 or less; (2) resolution of PTA by 3 months post‐injury; and (3) aged between 18 and 75 years. For the current study, at least one MRI was required, and patients were excluded if there was a known intervening neurological event between MRIs (e.g., another TBI or intracranial infection; see Table 1 for patient details). For the full list of inclusion/exclusion criteria and further details of parent study, see Green et al. [20].

Table 1 shows the demographic and injury characteristics of the patient sample. Information is provided for all participants analyzed. All medical information was obtained through medical records and/or clinical interviews. Socio‐economic status classification was based on the Standard International Socio‐Economic Index of Occupational Status (using Ganzeboom updated tables: International Standard Classification of Occupation 2008) [22, 24] and PTA classifications were based on Orman et al. [23].

TABLE 1.

Injury and demographic characteristics of TBI sample (N = 123) and controls (N = 32).

Variable	TBI (N = 123)		Control (N = 32)		p
	Mean (SD, range)	Proportion	Mean (SD, range)	Proportion
Demographics
Age (years)	38.8 (15.6, 18–73)		41.5 (14.5, 27–56)		ns
Education (years)	13.9 (2.9, 4–21)		15.6 (2.8, 13–18)		0.0032
Estimated premorbid IQ (WTAR) (N = 112)	99.4 (17.9, 82–117)
Sex (male)		72.0%		47.0%	0.0067

Socio‐economic status (ISEI—SES category based on Ganzeboom et al. [ 22 ])
Very low (laborers, cleaners)		4.0%
Low (service workers, drivers)		33.0%
Lower‐middle (skilled trades, technicians)		24.0%
Upper‐middle (teachers, nurses, analysts)		24.0%
High (engineers, lawyers, executives)		15.0%
(N.B. For SES Category, N = 114 as the ISEI was not calculated for eight individuals who were students and one individual who was a housewife at the time of injury.)
Type of injury
Motor vehicle accident		58.5%
Fall		35.0%
Assault		3.3%
Sports injury		2.4%
Unknown/other		0.8%
Severity of injury
GCS (lowest of recorded scores)		6.3 (3.6, 3–15)
Mild (13–15)		10.6%
Moderate (9–12)		4.9%
Severe (≤ 8)		67.4%
Missing data		17.1%
Length of post‐traumatic amnesia (classifications based on Orman et al. [23])
Mild (≤ 1 day)		3.3%
Moderate (> 1 and < 7 days)		16.1%
Severe (> 7 days)		70.0%
Missing data		10.6%

Abbreviation: ISEI, International Socio‐Economic Index of occcupational status.

2.1.2. Healthy Control Group

N = 32 healthy control participants completed one or two MRI scans. These individuals were students and hospital staff members or friends or family members of students and staff (see Table 1). Key inclusion criteria for the larger study were aged 17 years or older and willingness to complete two MRI scans. Key exclusion criteria were a history of TBI or concussion or any disease/disorder primarily or secondarily affecting the central nervous system (e.g., MS, lupus). Eligibility was determined through an adapted, structured screening interview [25].

2.2. Design and Procedures

Using a prospective, repeated measures design, MRIs were undertaken at one or more time windows. The first was completed for all patients at a mean of 154 days post‐injury (5.1 months; n = 121, SD = 31 days, range = 112–311 days post‐injury). The second scan was completed at a mean of 406 days post‐injury (1.1 years; n = 79, SD = 68 days, range = 329–681 days post‐injury). The third scan was completed at least 2 years post‐injury, with a mean of 1024 days post‐injury (2.84 years; n = 79, SD = 352 days, range = 730–2442 days post‐injury). Please note, for simplicity of exposition our labels for these MRI time points are rounded to the nearest month or year (i.e., 5 months, 1 year, and 3 years). We had eight participants who had a fourth scan at a mean of 1498 days post‐injury (4.16 years; SD = 470 days, range = 1133–2504 days post‐injury) and in order to maximize the power of the statistical models, which benefit from the widest range of time points possible, we included their data. For control participants, 32 participants completed the first scan and 20 of these completed a second scan. The mean gap between scans was 432 days (1.2 years; SD = 270 days, range = 231–1182 days).

The MRI outcome variables were as follows: VBR, total volume of the left hippocampus (HPC‐L) and substructures (i.e., HPC‐L head, HPC‐L body, HPC‐L tail), total volume of the right hippocampus (HPC‐R) and substructures (i.e., HPC‐R head, HPC‐R body, HPC‐R tail), and total volume of the CC and substructures (i.e., CC genu, CC body, CC splenium).

MRI scans were acquired on a General Electric (GE) Signa‐Echospeed 1.5 Tesla HD scanner (SIGNA EXCITE, GE Healthcare, Milwaukee, WI, USA), using an eight channel head coil. Please see Supporting Information document—Text 1 (pp. 2–6) for full details of acquisition, preprocessing, and manual segmentation steps.

2.3. Statistical Analysis

Mixed‐effects longitudinal modeling was used in the analyses, in which the change in volumes over time is treated as a random effect that varies from subject to subject. The models use splines for the post‐injury volume trajectories to allow for the possibility that the rate of change differs in earlier versus later stages of recovery. Thus, the models allow not only the estimation of average changes in volumes but also the estimation of between‐patient variability in the net underlying rates of recovery, above and beyond variability attributable to measurement error.

The first study objective was achieved through the use of splines with different rates in the earlier and later stages of recovery. The second objective was achieved via random effects, whose variability allows estimation of the proportion of patients whose rates of decline fall within various intervals/stratifications.

A summary of the rationale for the use of mixed‐effects longitudinal data for the current dataset can be found in Supporting Information—Text 2 (pp. 7–8).

Dependent variables consist of volumetric measures and VBR. These measures are by nature positive, with typical sizes varying from structure to structure, and for the same structures, from person to person. Thus, analysis of raw measures produces coefficients that may be difficult to interpret and compare. A change of, say, 10 voxels in a particular structure has a different significance if it occurs for a person for whom that structure is large compared with a person for whom it is small.

In addition to analyzing raw volumetric measures, we also perform analyses using a log transformation (natural logarithm × 100) for each volumetric measure as a dependent variable. For Y = a volumetric measurement, and $Y^{\prime }$ = the log transformation, we get:

$$Y^{\text{'}}=100\times \ln Y$$

With $Y^{\prime }$ as a dependent variable, the regression coefficients represent the approximate percent change in the expected value of $Y$ associated with a unit change in a predictor variable. The percent change thus estimated is the instantaneous rate of change, which is continuously compounded to yield the proportional rate of change over an interval of time. If the coefficient has a relatively small value (e.g., between −10 and 10), the associated percent change associated with a unit change in the predictor is close to the parameter of the regression.

With mixed models, the variability in trajectories between patients is an explicit component of the model so that the variability in the population can be estimated from the fitted parameters of the model. Using the log‐transformed dependent variables allows direct estimation of the proportion of a patient population similar to that analyzed in this sample that would exhibit various degrees of percent change in volumetric measure both in the short and in the longer term.

The equation defining the magenta line within the first year post‐injury, that is, for $32<\mathrm{ag}{\mathrm{e}}_{\text{patient}}<33$, is:

$$E\left(Y\right)={\gamma }_0^p+{\gamma }_1^p\left({\mathrm{age}}_{\text{patient}}-30\right)+{\gamma }_2^p\left({\mathrm{age}}_{\text{patient}}-32\right)$$

Thus the rate of change per year in this period is ${\gamma }_1^p+{\gamma }_2^p$.

After 1 year post‐injury, that is, for $33<\mathrm{ag}{\mathrm{e}}_{\text{patient}}$, the equation defining the magenta line is:

$$E\left(Y\right)={\gamma }_0^p+{\gamma }_1^p\left({\mathrm{age}}_{\text{patient}}-30\right)+{\gamma }_2^p\left({\mathrm{age}}_{\text{patient}}-32\right)+{\gamma }_3^p\left({\mathrm{age}}_{\text{patient}}-33\right)$$

The rate of change per year in this period is ${\gamma }_1^p+{\gamma }_2^p+{\gamma }_3^p$.

The estimated line for controls is ${\gamma }_0^c+{\gamma }_1^c\left({\mathrm{age}}_{\text{control}}-30\right)$, whose slope with respect to age is ${\gamma }_1^c$. Thus, the estimated level and slope for patients can be compared with the level and slope of this line.

2.3.1. General Formulation of the Model

We sought to estimate both the magnitude and the rate of change of the gap between patients and controls during the first 2 years post‐injury, controlling for injury age and concomitant aging effects post‐injury. If patients and controls come from equivalent populations, this gap represents the average effect of injury on the volumetric trajectory. The size of the gap at various times post‐injury provides an estimate of the effect of injury. A widening of the gap corresponds to patient deterioration while a narrowing represents improvement (Figure 1).

FIGURE 1.

Visualizing the model. The magenta line represents an estimated average trajectory for patients who sustain an injury at age 32. Separate slopes are estimated for the period before 1 year post‐injury and after 1 year post‐injury. The black line represents the estimated level for uninjured controls. The distance between the black control line and the magenta line is the estimated gap between the mean patient level and the level for uninjured controls of the same age. A widening gap reflects a rate of deterioration for patients that is greater than that of controls of the same age. A narrowing gap reflects mean recovery. If the magenta line segment is parallel to the black control line, the relative rate of change for an average patient is equal to the average of controls.

The model for controls uses a linear effect of aging through the period from age 20 to age 65, a model that is consistent with our data. The model for each response variable among controls is as follows:

$$Y_{\mathit{it}}={\gamma }_0^c+{\gamma }_1^c{\mathrm{age}}_{\mathit{it}}+u_i^c+e_{\mathit{it}}^c$$

where $\mathrm{ag}{\mathrm{e}}_{\mathit{it}}$ is the age (in years to the nearest day, centered at 30 years of age) of control subject $i$ on occasion $t$, ${\gamma }_0^c$ is the population intercept, $y_1^c$ is the predicted average change in $Y$ associated with aging 1 year, $u_i^c$ is the deviation in the random intercept for subject $i$ from the population intercept, and $e_{\mathit{it}}^c$ is the random variation in $Y$ at for subject $i$ at occasion $t$.

The random terms $u_i^c$ and $e_{\mathit{it}}^c$ are assumed to be independent and normally distributed with variances $g_{00}^c$ and ${\sigma }^2$, respectively.

The model for patients combines aging effects with a post‐injury recovery trajectory based on the elapsed time from the date of injury to the assessment date. The model allows for the possibility that the post‐injury recovery trajectory is not linear by fitting a linear spline with a knot at 1 year (see the gsp function in Monette et al.) [26]. This fits lines with possibly different slopes before and after 1 year with the constraint that the lines meet at 1 year. Fitting a spline makes it possible to test for a possible change in the direction and rate of recovery in the earlier versus the later stages of recovery. The model for a response variable, $Y$ without interaction between age and post‐injury trajectory, has the form:

$$Y_{\mathit{it}}={\gamma }_0^p+{\gamma }_1^p{\mathrm{age}}_{\mathit{it}}+{\gamma }_2^p{\text{post}}_{\mathit{it}}+{\gamma }_3^p{\text{post}}_{\mathit{it}}^{+}+u_{0i}^p+u_{2i}^p{\text{post}}_{\mathit{it}}+u_{3i}^p{\text{post}}_{\mathit{it}}^{+}+e_{\mathit{it}}^p$$

where $\mathrm{pos}{\mathrm{t}}_{\mathit{it}}$ is the number of years (to the nearest day) post‐injury for patient $i$ at measurement $t$, $\mathrm{pos}{\mathrm{t}}_{\mathit{it}}^{+}$ is the number of years beyond 1 year post‐injury, that is,

$${\text{post}}_{\mathit{it}}^{+}=\left\{0,\text{if}{\text{post}}_{\mathit{it}}\le 1{\text{post}}_{\mathit{it}}-1,\text{if}{\text{post}}_{\mathit{it}}>1\right.$$

With this form of the model, ${\gamma }_2^p$ is the population average rate of change in $Y$ during the first year in addition to that predicted by aging, ${\gamma }_3^p$ is the change in this rate of change between the second year and first year. Thus, the rate of change during the second year is ${\gamma }_2^p+{\gamma }_3^p$.

The random effects $u_{0i}^p$, $u_{2i}^p$, and $u_{3i}^p$ have a multivariate normal distribution with mean 0, varying from patient to patient, with variance matrix

$$G^p=\left(g_{00}^p{g}_{01}^p{g}_{02}^p{g}_{10}^p{g}_{11}^p{g}_{12}^p{g}_{20}^p{g}_{21}^p{g}_{22}^p\right)$$

These random effects are assumed to be independent of the error term for individual measurements, $e_{\mathit{it}}^p$, which has a normal distribution with mean 0 and standard deviation $\sigma$.

The gap between the expected value of $\gamma$ for a patient compared with a control subject at a given age and time post‐injury is estimated as:

$${\gamma }_0^p+{\gamma }_1^p\mathrm{age}+{\gamma }_2^p\text{post}+{\gamma }_3^p{\text{post}}^{+}-{\gamma }_0^c-{\gamma }_1^c\mathrm{age}$$

The rate of change in the size of the gap is:

$${\gamma }_1^p+{\gamma }_2^p-{\gamma }_1^c$$

before 1 year post‐injury and

$${\gamma }_1^p+{\gamma }_2^p+{\gamma }_3^p-{\gamma }_1^c$$

after 1 year post‐injury.

The model is extended to allow for possible interactions between age and the shape of the post‐injury trajectory and to allow for interaction with other variables such as proxies for the severity of injury. A detailed description of the model can be found in Supporting Information—Text 3 (pp. 9–10).

3. Results

In preliminary testing of the model, the following moderators were included: effect of age on post‐injury trajectory, years of education (YOE), and severity of injury (i.e., Glasgow Coma Scale and length of PTA). Preliminary analyses showed low moderator effects of YOE and severity of injury on trajectories. Therefore, they are not reflected in the following models. Many responses show a moderating effect of age on the trajectory slopes and the control‐patient gap, which is included in the models below.

Table S1 summarizes the Wald statistics for the base model trajectories (i.e., difference in percentage change among patients and percentage change among controls per year for Window 1 [5 months to 1 year post‐injury] and Window 2 [1–3 years post‐injury]) for VBR, HPC‐R, and HPC‐L and their substructures, and the CC and its substructures. These findings relate to our first objective, focusing on rates of change at the earlier versus later time windows (i.e., attenuation vs. acceleration). Here, we examined change in the rate of change (i.e., 5 months to 1 year and 1–3 years post‐injury) for each structure.

The models provide considerable evidence that there are differences between the slopes of patients and the slopes of controls, for both Windows 1 and 2 for VBR and CC, and for the majority of HPC‐R and HPC‐L substructures. In other words, there is evidence of greater volume loss in these structures and substructures in patients compared to controls, as observed in our prior research [5, 6, 20]. By including a linear spline in the model with a knot at 1 year post‐injury, the models also offer evidence that there are changes in these relative slopes (i.e., changes in the rate of change), in the direction of attenuation, between the period prior to 1 year and after 1 year post‐injury (Objective 1).

Table 2 supplements the previous findings by providing estimated differences in absolute total volume (mm³) between patients and controls for HPC‐L, HPC‐R, and the CC for selected ages at time of injury and for selected times post‐injury.

TABLE 2.

Estimates of mean patient and control volume differences at selected ages at time of injury (i.e., 20, 40, and 60) and for selected times post‐injury (5 months, 1 year, 3 years post‐injury) for left and right HPC and CC.

Region of interest	Patient age	Time point	Estimate (# voxels)	SE	t‐value	p
CC total	20 years old	5 months	−106.3	42.6	−2.5	0.014
		12 months	−151.1	41.9	−3.6	0.00047
		3 years	−178.9	40.8	−4.4	0.00003
	40 years old	5 months	−63.9	25.7	−2.5	0.014
		12 months	−108.6	25.7	−4.2	0.00005
		3 years	−136.5	25.8	−5.3	< 0.00001
	60 years old	5 months	−21.4	43.8	−0.49	0.63
		12 months	−66.2	44.5	−1.5	0.14
		3 years	−94.1	45.6	−2.1	0.042
HPC‐L total	20 years old	5 months	−360.7	183.2	−2.0	0.051
		12 months	−460.8	182.8	−2.5	0.013
		3 years	−485.8	182.7	−2.7	0.0090
	40 years old	5 months	−472.9	154.3	−3.1	0.0027
		12 months	−573.1	154.2	−3.7	0.00032
		3 years	−598.0	154.9	−3.9	0.00019
	60 years old	5 months	−585.2	182.7	−3.2	0.0018
		12 months	−685.3	183.0	−3.7	0.00029
		3 years	−710.3	184.3	−3.9	0.00019
HPC‐R total	20 years old	5 months	−244.9	147.0	−1.7	0.098
		12 months	−369.5	147.4	−2.5	0.014
		3 years	−420.9	146.6	−2.9	0.0050
	40 years old	5 months	−261.2	115.7	−2.3	0.026
		12 months	−385.8	116.9	−3.3	0.0013
		3 years	−437.2	117.1	−3.7	0.0003
	60 years old	5 months	−277.5	146.9	−1.9	0.061
		12 months	−402.1	148.3	−2.7	0.0078
		3 years	−453.5	149.5	−3.0	0.0030

Note: VBR is a ratio and does not permit this quantification.

Abbreviations: CC, corpus callosum; HPC, hippocampus.

Table 3 displays the distribution of the random effects portion of the model that allows us to estimate the between‐patient variability in relative slopes and, thus, to estimate the proportion of patients whose relative slopes fall in selected windows, assuming an approximate normal distribution of relative slopes (Objective 2). Here, the estimated annualized rate of decline in patients was adjusted for the estimated annualized aging‐related atrophy estimated from healthy control participants.

TABLE 3.

Estimated proportion of patients showing decline relative to healthy control participants in VBR, CC, and HPC and their substructures stratified by rates of decline.

Outcome	Time window	Mean	SD	Percentage of patients with atrophy ≥ 2.5%	Percentage of patients with atrophy within selected subintervals
					0%–2.5% atrophy	2.5%–5% atrophy	5%–10% atrophy	> 10% atrophy
VBR	Window 1 (5–12 months)	20.3	14.5	89.1	2.9	3.6	9.3	76.2
	Window 2 (12 to 3 years)	6.1	3.2	87.0	10.2	23.7	52.4	10.9
CC total	Window 1	10.3	10.3	79.0	4.9	7.6	18.2	53.2
	Window 2	4.6	4.7	67.1	9.9	20.8	34.1	12.2
CC genu	Window 1	9.4	16.4	66.2	4.9	5.7	12.0	48.5
	Window 2	4.3	5.1	63.8	10.9	19.2	31.3	13.3
CC body	Window 1	12.5	11.4	80.9	4.2	6.5	15.7	58.7
	Window 2	5.2	3.4	78.8	5.1	26.1	44.6	8.1
CC splenium	Window 1	11.3	12.9	75.3	4.8	6.5	14.7	54.1
	Window 2	4.6	9.0	59.3	9.0	11.0	20.9	27.4
HPC‐L total	Window 1	4.6	6.8	62.1	10.2	14.6	26.4	21.1
	Window 2	0.7	2.8	26.4	27.3	20.1	6.3	0.0
HPC‐L head	Window 1	3.5	10.0	54.2	8.9	10.0	18.3	25.9
	Window 2	0.3	2.7	20.0	31.1	16.2	3.8	0.0
HPC‐L body	Window 1	4.6	11.6	57.2	7.6	8.6	16.5	32.1
	Window 2	2.0	5.2	46.5	15.6	18.0	22.2	6.3
HPC‐L tail	Window 1	8.7	11.7	70.4	5.9	7.8	16.9	45.7
	Window 2	0.9	11.1	44.1	8.8	8.7	14.9	20.5
HPC‐R total	Window 1	5.6	7.9	65.1	8.7	12.2	24.1	28.8
	Window 2	1.4	3.9	39.0	20.1	21.3	16.4	1.3
HPC‐R head	Window 1	4.4	9.5	57.8	8.8	10.5	19.7	27.6
	Window 2	1.8	4.2	43.3	18.1	21.0	19.7	2.6
HPC‐R body	Window 1	5.9	9.1	64.6	8.1	10.7	21.4	32.5
	Window 2	2.8	6.8	51.5	12.3	14.4	22.8	14.3
HPC‐R tail	Window 1	9.9	19.2	65.1	4.4	4.9	10.3	49.9
	Window 2	1.3	7.5	30.7	13.3	10.6	13.5	6.6

Note: Distribution of the random effects portion of the model estimating between‐patient variability in relative slopes. Note that the estimated annualized rate of decline in patients was adjusted for the estimated annualized aging‐related decline estimated from healthy control participants.

Abbreviations: CC, corpus callosum; HPC, hippocampus; VBR, ventricle‐to‐brain ratio.

As can be seen in Table 3, well over half of the patients are estimated to have annualized relative rates of atrophy of greater than 2.5% compared to healthy controls for VBR and CC total at both the early and later chronic stages. It is estimated that a smaller proportion of patients declined in the later chronic stages, and rates of decline were less steep. More specifically, in Window 1 (i.e., 5 months to 1‐year post‐injury) and Window 2 (i.e., 1–3 years post‐injury), 89.1% and 87.0%, respectively, showed an estimated rate of atrophy of at least 2.5% for VBR. From 5 months to 1 year post‐injury, the majority of patients (76.2%) are estimated to fall in the fastest rate of atrophy interval (> 10%) whereas from 1 to 3 years post‐injury, the majority (52.4%) are estimated to fall in the middle interval (5%–10%). For CC total, from 5 months to 1 year post‐injury, 79.0% of patients are estimated to have a rate of loss of at least 2.5% greater than controls, with 53.2% falling in the highest rate of change interval (> 10%). From 1 to 3 years post‐injury, 67.1% are estimated to continue to show a rate of atrophy of at least 2.5%, 34.1% fall in the 5%–10% rate of change interval, 20.8% in the 2.5%–5% range, and 12.2% in the > 10% interval. The substructures of the CC follow a similar pattern.

Patterns for HPC were different with a somewhat lower percentage of patients estimated to show progressive decline in the early time window and a lower percentage predicted to show progressive losses in the highest rate of change interval (> 10%). For total volumes of the HPC‐L and HPC‐R, approximately two‐thirds of patients are estimated to show atrophy of more than 2.5% from 5 months to 1 year post‐injury. In this early window, for HPC‐L, 21.1% of patients are predicted to fall in the fastest annual rate of change interval (> 10%) and 26.4% are predicted to show an annual rate of change in the middle interval (5%–10%). Similarly, for HPC‐R 28.8% of patients are predicted to show an annual rate of atrophy > 10%, while 24.1% fall in the middle interval (5%–10%). From 1 to 3 years post‐injury for HPC‐L and HPC‐R, the greatest percentage of patients is predicted to show a rate of atrophy in the 0%–2.5% and 2.5%–5% range. The substructures of the HPC show a similar pattern.

4. Discussion

The current findings shed light on an understudied area of research: the deteriorative changes that occur within the chronic stages of msTBI. Addressing the question whether rate of atrophy accelerates, abates, or remains constant over time, our models were encouraging. Despite estimating robust and widespread declines at the first window and ongoing decline at the second, there was nonetheless a significant tapering of decline between the windows. Asking what proportion of patients showed progressive volume loss in the earlier and later time windows, and the magnitude of that atrophy, a large number of patients were estimated to show extensive progressive atrophy across all ROIs. VBR showed the greatest proportion of patients with progressive declines, and the greatest proportion at the highest magnitude of loss (> 10%). Here, there was notably greater estimated loss in the first window (76.2% with > 10% annualized loss) than the second window (10.9% with > 10% annualized loss); nonetheless, more than 50% of patients still showed an annualized decline in the second window of 5%–10%. Encouragingly, the CC showed the same pattern of tapering, but milder findings overall, and for the HPC, the proportion and severity of progressive decline were milder still. Nonetheless, even at the second time window, the left and right HPC were still showing an estimated 26% and 39% of patients, respectively, with progressive atrophy of at least 2.5% as compared to controls. To our knowledge, this is the first study to model chronic msTBI data with this nature and degree of detail concerning the timing, regional variation, proportion, and severity of progressive volumetric declines.

In our previous research examining an earlier subset of these patients and a smaller subset of controls, we reported significant monthly percent decline of the same regions as compared to controls from 5 to approximately 20 months post‐injury [20]. In Belchev [6], we examined the same time windows as in the current study with different (but partially overlapping) regions; we employed automated segmentation rather than manual. We found that from 5 months to 1 year, patients showed significantly increased atrophy in the accumbens and marginal atrophy in the amygdala, brainstem, thalamus, and the left HPC followed by continued progression at the same rate—without attenuation. These differences between the previous and current findings may be explained by the different structures assessed, lesser accuracy of automatic segmentation, and/or differences in analytic approaches. (The former study employed exploratory principal components analysis followed by an examination of nonlinear trends in the early vs. later time windows with patients only; the latter employed mixed effects growth curve modeling, which harnessed all data rather than exclusively those with all three time‐points and controlled for aging‐related effects throughout.) Further research is needed that combines whole‐brain measures plus a range of cortical and subcortical structures within the same study and that harnesses the strengths of manual segmentation and mixed effects growth curve modeling—which allows for not only estimation of average changes in volumes, but also between‐patient variability in the underlying rates of recovery.

With regard to research in other groups, a handful of longitudinal studies with controls have examined the whole brain, CC, and HPC [12, 13, 27, 28, 29] using varying time windows. Employing similar time windows, Sidaros et al. [12] reported −4.0% volume loss in the whole brain in patients versus −0.18% in controls between 2 and 12 months post‐injury, while Cole et al. [27] reported −1.5% versus −0.22% from 1 to 2 years post‐injury. In contrast, longitudinal studies examining the CC and HPC showed somewhat greater volumetric reductions [28]. Again, differences from our findings, particularly in studies that used similar time windows, may reflect the more fine‐grained examinations of the data in the current study, which stratified the proportion and degree of atrophy (plus manual segmentation and longitudinal mixed effects modeling).

With regard to the cause of these declines, the MRI quantitation cannot distinguish between volume losses attributable to cell death versus shrinkage of the cell body (e.g., decreased proliferation of proteins) or loss of neuropil [30]. Nonetheless, the ongoing and substantive estimates of volume loss from 1 to 3 years provide strong evidence that losses cannot be explained simply (and more benignly) by the resolution of acute injuries, for example, resolution of edema or Wallerian degeneration [31, 32]. Likely, the drivers of chronic neurodegeneration are multifactorial and synergistic, with effects on white matter adversely affecting gray matter and vice versa. Environmental factors may play a deleterious and synergistic role as well. After msTBI, patients frequently endure reduced schedules of activity [33]. They may have reduced physical activity (e.g., due to ongoing physical disability and reduced access to gyms) as well as reduced cognitive and social stimulation because of unemployment and diminished social networks [34]. Although “use” is neuroprotective in aging and neurological disorders [33, 34, 35], the absence of these factors after msTBI has been postulated to give rise to negative neuroplastic changes or disuse‐mediated atrophy [34, 35]. The more tempered losses observed in the left HPC raise the question whether the informal day‐to‐day activities of patients (conversations; watching television) confer ongoing stimulation/use specifically to this verbal learning structure, and thereby neuroprotection. Neurogenesis of the HPC may also explain the relatively lesser volume losses [36]. Moreover, brain damage itself causes disconnections resulting in lost projections to areas originally unaffected by acute brain damage (anterograde trans‐neuronal degeneration); this reduced connectivity and commensurate loss of neurotrophic factors may again contribute to disuse‐mediated neural deterioration over time [34]. One might expect trans‐neuronal degeneration caused by such disuse to be self‐limiting, which may explain the reduced atrophy across the brain in the later time window.

At the molecular level, a number of mechanisms may explain changes to the brain in the chronic stages. These include protracted neuroinflammation of the white matter with damage to microtubules and reduced axonal transport (with Wallerian‐like degeneration), and associated microglial activation [37, 38, 39, 40]; microglial activation has been observed > 15 years following a single TBI [41]. Propagation of harmful proteins along damaged white matter networks associated with protein misfolding is another implicated mechanism [42]. Ongoing blood–brain barrier disruptions, which themselves are associated with neuroinflammation and anomalous protein deposition, delayed apoptosis [43], damaged glymphatic system [44], metabolic overuse/dysregulation [45, 46, 47], and dampened neurogenesis [48], are all mechanisms which have been implicated in the chronic stages of msTBI.

4.1. Clinical Implications

We contend that clinicians prognosticate outcomes after msTBI under the presumption of stability of the brain into the chronic stages of injury [49]. Most clinical care—particularly in contexts where care is supported by public health systems (e.g., Canada, United Kingdom, Australia)—is frontloaded in the early weeks and months of injury when spontaneous recovery is maximal [50, 51]. Unfortunately, there is little ongoing monitoring or treatment in the chronic stages, and the absence of intervention combined with reduced schedules of activity may exacerbate decline [33, 34]. These results underscore the need for ongoing treatment into the chronic stages, particularly for patients at the more severe end of the injury spectrum. Although we observed attenuation of rate of neural decline, our past cognitive findings [14, 18] showed an increasing proportion of patients demonstrating cognitive declines from 2 to 5 to 12 to 24+ months post‐injury, with progressive HPC loss (left side) from 5 to 12 months post‐injury a predictor of proportion of decliners from 12 months to 2+ years post‐injury. We suggest that cumulative effects of ongoing brain volume loss give rise to this seeming contradiction. This again highlights the need for ongoing treatment ‐ and research ‐ into the chronic stages of injury.

Another clinical implication of this line of research concerns the relationship between biomarkers and the development of disease‐modifying therapies. Progressive brain volume loss is emerging as a useful biomarker in other neurological disorders and becoming advocated for TBI [52, 53]. This research advances our understanding of the expected rates of deterioration, by region, and at varying time windows after injury, aiding in the utility of brain volume loss as a brain biomarker for msTBI research.

4.2. Limitations

Limitations of the current study include the limited sample size, particularly combined with the heterogeneity of patients in terms of neuropathology and pre‐injury variabilities (a challenge of all TBI research). A sampling limitation was also the homogeneity of injury severity, which may have reduced its predictive validity. The sample skewed toward very severe based on PTA (though not GCS, a discrepancy reported in the literature) [54], which limits the generalizability of the findings. There were some changes to the neuroimaging platform during the course of the study, which may have changed (enhanced) sensitivity of the imaging, but would arguably not have introduced biases as modeling is able to accommodate increased heteroskedasticity associated with moving from lower to higher accuracy of measurement. Unresolved edema in the first window of assessment is an improbable, but potential confound. The presence of edema would inflate volume at the first assessment, creating a steeper volumetric decline across the first two assessments. Lastly, the work was undertaken using 1.5 T MRI. Arguably, the strong signal despite the lesser sensitivity suggests that the current findings might have been even more pronounced at greater magnet strength.

4.3. Conclusions

This is the first study to quantify the proportion of patients showing progressive atrophy across two chronic windows of msTBI, to provide evidence of attenuation of atrophy across time, and to estimate the rates/degree of decline in different time windows. These findings of robust and ubiquitous progressive atrophy occurring in the early months and years following msTBI (albeit to a lesser extent in the later stage measured in this study), underscore the scale of the problem. Research is needed to understand the behavioral implications of these findings, to evaluate how patterns may vary across different brain regions and a broader range of injury severity within the moderate to very severe range, and to identify strategies for prevention of declines.

Author Contributions

Conception and design of the study: R.E.A.G., B.C., and D.J.M. Acquisition and analysis of data: R.E.A.G., B.C., D.J.M., K.J., and G.A.M. Drafting the manuscript or figures: R.E.A.G., B.C., A.C., A.L., and G.A.M. Study supervision: R.E.A.G.

Funding

This work was supported by Canadian Institutes of Health Research (MOP 86704), Physicians' Services Incorporated Foundation (12‐43), and the Canada Research Chairs Program.

Ethics Statement

The Research Ethics Board at University Health Network, Toronto, has approved the use of humans for this study and the study was completed in accordance with the ethical standards outlined by the research ethics board.

Consent

Written informed consent was obtained from each study participant. No identifying information has been included in this article.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting Text 1: MRI Acquisition and processing; Strengths of the multi‐level modelling approach for the current dataset; Details of model.

Supplementary Table S1: Patient Annualized Percent Rate of Change Minus Control Annualized Rate of Change 5 to 12 months and 1 to 2+ years.

Funding Statement

This work was funded by Canadian Institutes of Health Research grant MOP 86704; Physicians' Services Incorporated Foundation grant 12‐43; Canada Research Chairs Program ; Walter and Maria Schroeder Foundation ; Joseph & Antoinette Sorbara Foundation .

Data Availability Statement

The principal author takes full responsibility for the data, the analyses and interpretation, and the conduct of the research. The principal author has full access and the right to publish all data. Investigators may request access to the anonymized patient data that support the findings of this study from the principal author. Prior to use of the data, proposals need to be approved and a signed data sharing agreement must be completed.