Accelerated Brain Aging in Mild Traumatic Brain Injury: Longitudinal Pattern Recognition With White Matter Integrity

Background: Long-term effects of mild traumatic brain injury (mTBI) resemble brain aging changes (i.e., microstructure integrity loss), which implies an accelerated age-associated process. This study aimed to develop a quantiable neuroimaging marker to characterize the brain-aging process accelerated by mTBI from acute to chronic phases. Methods: A brain-age prediction model was dened using relevance vector regression (RVR) in 523 healthy individuals, based on fractional anisotropy metrics from diffusion-tensor imaging. The model was adopted to estimate brain-predicted age difference (brain-PAD = predicted brain age - chronological age) in 116 acute mTBI patients and 63 healthy controls (HCs). Fifty patients were followed up 6~12 month post-injury to evaluate the longitudinal changes in brain-PAD. Another mTBI group containing 70 acute patients were included as a replicated cohort. We investigated whether brain-PAD was greater in patients with elderly age, post-concussion complaints, and risky apolipoprotein E (APOE) genotype, and whether it had the potential to predict neuropsychological outcomes for information processing speed (IPS). Between-group and longitudinal comparison in brain-PAD was conducted with analysis of covariance and linear mixed-effects model, respectively. The correlation between brain-PAD and continuous variables was analyzed with Spearman rank-order correlation. Results: The RVR brain-age prediction model predicted brain age accurately (r = 0.96, R 2 = 0.93). The brain age of mTBI patients was estimated to be "older" in the acute phase, with mean brain-PAD of 2.59 (± 5.97) years compared with HCs (0.12 ± 3.19 years) (P < 0.05) and replicated in another mTBI cohort (brain-PAD: 3.26 ± 4.55 years). The increased brain age in mTBI kept stable at 6-12 month post-injury (2.50 ± 4.54 years). Patients with older age or severer post-concussion complaints obtained greater brain-PAD (P < 0.001, P = 0.024), while patients with APOE ε4 didn’t obtain greater brain-PAD than those without. Additionally, brain-PAD in the acute phase predicted patients’ IPS prole at 6~12 month follow-up (rho = -0.36, P = 0.01). Conclusion: Mild TBI, even a single one, accelerates the brain-aging process. The brain-PAD can be considered as a quantitative neuroimaging marker to evaluate the susceptibility to neurodegeneration or other age-associated conditions following mTBI. (RVR) input 10-fold cross-validation on the training set with the same error calculation of the test set.


Introduction
Mild traumatic brain injury (TBI), accounting for 80 ~ 90% of the total TBI [1], is typically associated with microstructural integrity loss that persists chronically after the initial injury [2]. The long-term sequelae of mild TBI (mTBI) initiate the risks for neurodegeneration and other age-associated conditions, making it a clinically important topic. Recent large-population retrospective and observational studies showed that chronic mTBI associates with a higher risk of Parkinson's disease (hazard ratio: 1.56) and dementia (hazard ratio: 1.17) [3,4]. However, neuroimaging biomarkers that evaluate the individual-based high risks of age-associated or neurodegenerative deterioration in the brain for mTBI patients are still scarce.
An increasing number of researches have demonstrated that MRI metrics strongly relate to chronological age [5,6]. The deviation between chronological age and predicted brain age can re ect how brain diseases interact with normal aging [7][8][9][10] and predict the risk of neurodegenerative diseases [11]. The increases in brain-predicted age difference (brain-PAD = predicted brain age -chronological age) have been widely presented in psychiatric disorders and Alzheimer's disease [12,13]. Recently, Cole et al. observed a 4.66-year increase in brain-PAD estimated by morphometrics in white matter and a 5.97-year increase estimated by gray matter morphometrics in chronic moderate/severe TBI patients instead of the mild cases [10]. However, these results limited our understandings of how mTBI accelerated the brainaging process because only 17 patients with mTBI were included in their research. The majority of TBIs sustained in civilian contexts are within the mild range [14], which is associated with increased risks of dementia in older adults [15]. The diffuse axonal injury lead by mTBI [16] causes the long-term deterioration in white matter tracts (WMTs) integrity [17], which is mainly responsible for remote neurodegeneration and aging-related conditions [18,19]. Therefore, we hypothesized that mTBI might accelerate the aging process by loss of microstructural integrity.
The relevance vector regression (RVR) [20] was applied to build a brain-age prediction model in 523 healthy individuals. This model was then used to estimate brain age in 116 acute mTBI patients and 63 healthy controls (HCs). Fifty patients were followed up 6 ~ 12 months to evaluate the longitudinal changes in brain-PAD. Another independent 70 patients were included as a replicated cohort. We investigated whether brain-PAD was prone to be greater in patients with elderly age, post-concussion complaints, and risky apolipoprotein E (APOE) genotype [15,[21][22][23], and whether it can predict the neuropsychological outcome for information processing speed (IPS). In addition, we hypothesized that white matter tracts contributing to the increased brain age in mTBI patients were primarily located in the aging-degenerative regions.

Participants
Training set. Diffusion-tensor imaging (DTI) scans from 523 anonymized healthy individuals were used to build the brain-age prediction model (details in eMethods of Supplemental Materials and Supplemental Table e1).
Test sets. The The ow of participants through the study is shown in Figue 1. There were 116 acute mTBI patients recruited from the local emergency department between 2016 and 2018. Sixty-four patients consented to join in the follow-up investigations, and 50 of them were nally followed 6~12 month postinjury. Screening for mTBI was based on the World Health Organization's Collaborating Centre for Neurotrauma Task Force [24] (details in Table 1). The retrospective baseline health conditions in patients (i.e., hyperglycaemia, hypertension, and hyperlipidemia) were recorded at their rst visit. Healthy controls (HCs, N = 63) were enrolled through the advisements and carefully screened for any neurological or psychiatric disorders. In addition, an independent cohort including 70 acute mTBIs with different MRI scanner was included for replication use (details in eMethods of Supplemental Materials). All the subjects gave written, informed consent in person approved by the local institutional review board and conducted in accordance with the Declaration of Helsinki.

Clinical Symptom and Neuropsychological Assessment for Patients
Clinical symptoms were assessed at the rst visit of patients. The post-concussive symptoms (PCS) was evaluated by the International Classi cation of Diseases, 10th edition clinical (ICD-10) criteria [25].
Patients with three or more symptoms were classi ed as with PCS (PCS+) otherwise without (PCS-). According to the anterograde amnesia duration, patients were divided into with anterograde amnesia (AA+) and without (AA-).
The information processing speed (IPS) was assessed by the Digit Symbol Coding test [26] and Trail-Making Test (Part A) [27] at both initial and follow-up visits. Z-score composed from these two tests was summarized as the performance of IPS.

Neuroimaging Data Acquisition and Preprocessing
The training dataset was collected from multi-centers, by using different MRI scanners (Table e1 in Supplemental Materials). The intraclass correlation coe cient (ICC) of the homogeneity of datasets was calculated after preprocessing. The index ICC for fractional anisotropy (FA) was 0.84, showing the "excellent" homogeneity (ICC > 0.8) for multi-center datasets.
All raw DTI data were quality controlled (details in eMethods of Supplemental Materials) and preprocessed using FSL5.0.9 software package (http://www.fmrib.ox.ac.uk/fsl/index.html) following the standard procedure [28]. Fifty WMTs were selected based on rICBM_DTI_81_WMPM_FMRIB58 1mm atlas and the mean FA value in each WMT was calculated (details in eMethods of Supplemental Materials).

Brain Age Prediction Modeling Procedure
The brain-age prediction was performed by de ning a relevance vector regression (RVR) to model the relationship between FA values and chronological age in the training set, using Scikit-learn Python library.
AS a sort of probabilistic extended linear model based on Bayesian formulation, RVR is derived from Relevance Vector Machine (RVM). Concretely, in the current study, the predictive regression model for brain age was trained by estimating the relevance vectors and their corresponding weight distributions by using the type-II maximum likelihood algorithm. The relevance vectors were derived from the input data (i.e., FA values), which represented the prototypical examples associated with the regression target instead of solely separating attributes. The training process was visualized as Fig. 2A-2D (details in eMethods of Supplemental Materials). The prediction accuracy was assessed using the Pearson correlation (r) between chronological age and predicted brain age, the proportion of the variation (R 2 ) explained by the trained model, the mean absolute error (MAE), and root mean squared error (RMSE) of the brain-PAD scores.

Statistical Analysis
All the statistical analyses were conducted in the SPSS (version 21.0; IBM, Armonk, NY). To examine the increased brain age in test sets, we compared the brain-PAD scores with zero using the Wilcoxon Signedrank test. Mann-Whitney U test was used to examine the differences in PAD between the group with baseline health condition and the group without. Between-group comparison of the brain-PAD scores for mTBI samples and HCs was conducted by the Analysis of Covariance (ANCOVA), adjusting for age, gender, and education. The longitudinal comparison analysis was performed on brain-PAD between acute and chronic phase with the linear mixed-effects model (LME) analysis. Between-group comparisons in brain-PAD scores for PCS subgroups (PCS+ vs. PCS-), APOE genotypes (APOE ε4+ vs. APOE ε4-), AA subgroups (AA+ vs. AA-), and chronological age subgroups (youth: ≤ 30 years, middle-aged: 30~50 years, elderly: ≥ 50 years) were also conducted by the ANCOVA, adjusting for gender, education, and/or age. Spearman rank-order correlation (rho) was conducted to examine the associations of brain-PAD scores with education, post-injury time, and Z-score of IPS performance. For the mTBI cohort, the rho between brain-PAD and FA value of WMTs involved in model training was calculated to evaluate the related WMTs contributing to the increased brain age. Results were considered signi cant under P < .05 (two-tail). Effect sizes for ANCOVA and LME analysis were quanti ed using partial η 2 ; effect sizes for the Wilcoxon Signed-rank test and Mann-Whitney U test were quanti ed by the rank biserial correlation coe cient (r rb ).

Chronological age predicted by FA values
The RVR model could accurately predict the individual chronological age for both the training set and the HC test set. For the training set, age was accurately predicted by FA values (r = 0.96, R 2 = 0.93, MAE = 3.74, RMSE = 5.03). The mean PAD for the training group was -0.18 (± 5.03) years. For HC test set, brain age was also accurately predicted (r = 0.97, R 2 = 0.94, MAE = 2.57, RMSE = 3.16), and the mean PAD was 0.12 (± 3.19) years (details in eResults of Supplemental Materials).
In order to avoid the educational level and baseline health conditions of patients being confounding factors, we performed additional analysis. The educational years in the mTBI cohort were lower than that of HCs (P < 0.001). Therefore, we randomly selected a subgroup of mTBI patients (N = 108) and HCs (N = 47), who were matched with age, gender and education level, and then compared the difference in brain-PAD scores between groups. Acute mTBI patients still held higher brain-PAD scores (2.68 ± 5.99 years) than HC group (0.31 ± 3.27 years) (F 1, 173 = 6.60, P = 0.011, partial η 2 = 0.042). Additionally, the correlation analysis showed no statistically signi cant correlation between brain-PAD and years of education in mTBI patients (rho = -0.137, P = 0.143). Considering that baseline health conditions might affect the brain-PAD, we examine the differences between the subgroup of mTBI patients with a history of hyperglycaemia, hypertension, or/and hyperlipidemia (N = 8) and a matched group of patients who were randomly selected from whom without (N = 8). The result showed there was no statistical difference (W = 34, P = 0.88) in acute brain-PAD between two group patients. Those results showed the educational years and baseline health conditions of patients had no effects on PAD in the present study. Therefore, all 116 patients were included in the following analysis.
Neurodegeneration-vulnerable tracts associated with brain-PAD scores The rho between brain-PAD and FA value of WMTs involved in model training (N = 33) was calculated to evaluate the related WMTs contributing to the increased brain age. The FA value in 16 WMTs presented signi cantly correlated to the brain-PAD scores in the acute phase. They were mainly located in the commissural, association, and projection bers (rho: -0.305 ~ -0.473, P < 0.05 after family-wise error (FWE) correction for multiple comparisons) ( Table 3), primarily including the corpus callosum, fornix, corona radiate and superior longitudinal fasciculus. In the chronic phase, the FA value in the body of fornix was still signi cantly related to the brain-PAD at follow-up 6~12 months (rho = -0.434, P < 0.05, after FWE correction for multiple comparisons) ( Table 3).

Discussion
Our study revealed that mTBI triggered an accelerated aging process in the brain deviating from its chronological trajectory. The increased brain age was observed in the early acute phase and kept stable within a 6 ~ 12 month follow-up. The WMTs relevant to the accelerated aging-process were primarily located in the connecting aging-related and neurodegeneration-vulnerable regions. Besides, the brain-PAD of patients measured in the acute phase could predict the individual performance of IPS at 6 ~ 12 month follow-up. Patients with older chronological age and severer PCS complaints were more vulnerable to getting "older" brain age.
Unlike one recent study about increased brain-PAD observed in chronic patients (i.e., averaged over 49month post-injury) [29], the observation in our research covered a relatively early phase of mTBI. The results showed that the discrepancy between predicted brain age and chronological age (i.e., brain-PAD > 2 years) was detectable even within the 7-day post-injury. Moreover, the increased brain age cannot simply be explained as a transient physiological pattern in response to traumatic events. Instead, the current longitudinal observation on brain-PAD further demonstrated such an accelerated brain-aging process was a stable change. As described in neurological and neuropsychiatric researches, the increases in brain age have been considered as a surrogate neurobiomarker of vulnerability to neurodegeneration [7,30]. Likewise, the brain-PAD score may be termed as a candidate biomarker for mTBI. No signi cant correlations between the brain-PAD score and time since injury in both acute and chronic phases were observed, which further supplemented the explanation that such increases in brain age might re ect the vulnerability to the accelerated aging process.
For both acute phase and follow-up, the increases in brain age inversely related to the integrity of fornix. These evidences implied that the deteriorated integrity of fornix is mainly responsible for the persistent brain-aging process accelerated by mTBI. The fornix, connecting the hippocampal formation in the limbic system, is responsible for neurodegeneration progression [31]. Recent theory about the transmission model in Alzheimer's disease (AD) shows that the pattern spreads initially from the transentorhinal cortex and hippocampus latterly to the rest of the brain [32,33]. Moreover, alteration of microstructure organization in the fornix is also responsible for neurodegeneration in mild cognitive impairment and chronological aging process [34,35]. Those evidences in neurodegenerative conditions suggests that the fornix may serve as a common connecting origin of neurodegeneration in the transmission pathway.
Therefore, based on the current ndings on the steady association between the integrity of fornix and increased brain age, we can infer that brain-PAD score deriving from the WMT integrity was informative in evaluating the potential neurodegeneration effects.
We explored factors affecting brain-PAD in patients with mTBI, which could explain the individual differences in the acceleration of brain-aging post-injury. The results showed that patients with older chronological age and severer PCS complaints were more vulnerable to getting greater brain-PAD. These results were consistent with some retrospective studies' ndings that post-concussion syndrome and older age increased the risks of developing dementia for patients with mTBI [15,21]. The APOE genotypes or anterograde amnesia in our study showed non-signi cant effects on brain-PAD following mTBI. Both APOE ε4 and anterograde amnesia are generally considered as risk factors for AD [22,23]. Those provides some possible clues that the neurodegeneration-vulnerable aging process accelerated by mTBI may be differential from AD pathology [36]. However, there still call for further studies to explore the common and dissociated factors contributing to the neurodegeneration in mTBI and AD Our model explained the variations in the chronological aging process as aging-related changes in WMTs. Unlike previous studies using macroscopically morphological features to predict brain age in moderate to severe TBI patients [10,29], we adopted more subtle microstructural features, considering mTBI usually causes diffuse axon injury [16] and long-distance WM disconnections [18]. Those damages, thereafter, result in a degradation in the e ciency of communication among brain regions and contribute to cognitive dysfunctions during the brain-aging process, commonly re ected as slow speed of information processing and transferring [37]. Moreover, our results demonstrated that brain-PAD in the acute phase could predict the long-term individual pro les on the information processing speed (IPS). The greater brain-PAD was relevant to poorer performance on IPS, which suggests that WM integrity loss contributed to the cognitive impairments following mTBI. Additionally, the potential of WMTs as predictors for brain-PAD post-injury, discussed above, further held the fact that mTBI might be vulnerable to tracts primarily connecting the neurodegeneration-related regions.
There are limitations to the present study. Firstly, our brain-age prediction model was trained with DTI metrics with respect to injury characteristics of mTBI. The brain-aging process is biologically complex, which may be explained more by the multi-modality of neuroimaging. Secondly, in the current longitudinal study, we observed that the brain-PAD (i.e., 2.59 years) was stable within one-year, which re ected the aging-related changes persistently existing in white matter tracts after the injury. It still needs to be careful to infer whether or not this acceleration in brain aging process is responsible for the neurodegeneration, because the neurodegeneration following injury may contain a chronic interaction of neuroin ammation with damages to WMTs integrity [38]. Besides, the follow-up loss and insu cient APOE ε4 samples might also limit our understandings of the association between the accelerated brainaging process following mTBI and neurodegenerative diseases, such as AD.

Conclusions
In summary, the acceleration of the brain-aging process following mTBI is detectable even in the early acute phase and persists into the chronic phase, which provides insights into the risks of mTBI for neurodegeneration-vulnerable conditions. Moreover, the increases in brain age are related to clinical settings and capable of predicting neuropsychological outcomes. Abbreviations PAD = predicted age difference; mTBI = mild traumatic brain injury; FA = fractional anisotropy; WMT = white matter tract; RVR = relevance vector regression; DTI = diffusion-tensor imaging; IPS = information processing speed; PCS = post-concussive symptom; APOE = apolipoprotein E.

Declarations
Ethics approval and consent to participate Patient consent was Obtained. Ethics approval The Second A liated Hospital of Wenzhou Medical University.

Consent for publication
All the authors have approved the manuscript.

Availability of data and materials
All the data mentioned in this article are available on published article.

Competing interests
The authors have no con icts of interest to declare. Tables Table 1  The inclusion and exclusion criteria for mTBI   Inclusion and  exclusion criteria   Items   Inclusion     The ow diagram shows the inclusion process for participants of test sets. Abbreviations: brain-PAD = brain-predicted age difference; mTBI = mild traumatic brain injury; HCs = healthy controls.  Overview of brain age prediction model. (A) The model was performed in a given segment of the age span. (a) In the training set, the age span (11 ~87 years) was partitioned into three segments (phase 1: 11 32 years, N = 186; phase 2: 26 ~ 48 years, N = 156; phase 3: 42 ~87 years, N = 272). Two adjacent age phases in the training set were slightly overlapped (i.e., 6-year overlap) to enhance the tolerance towards the inter-individual variability of brain age prediction. (b) In the test sets, each age span segment corresponded with the age phase in the training set, in which the overlapped part of two adjacent phases was averaged (phase 1: 20 ~ 29 years; phase 2: 29 ~ 45 years; phase 3: over 45 years). (B) Feature selections contained two steps: (a) Pearson correlation coe cient (r) was calculated (P < 0.001) between the mean FA value of each WMT and age in each age segment of the training set, and WMT with the absolute value of r above 0.25 was kept; (b) Dimensionality of features was then reduced by Principle Components Analysis (PCA). The components with the top 90% variance were kept as the input features.
(C) The supervised learning process was conducted using Relevance vector regression (RVR) to de ne the relationship between the chronological ages and input features. Internal validation was assessed by running 10-fold cross-validation on the shu ed training set with the same error calculation of the test set. (D) The trained RVR brain-age model was entirely applied to all test sets.

Figure 3
MTBI increased the brain-aging process. (A) Scatterplot and linear regression of individual chronological age and brain-predicted age in each test set derived from the RVR prediction model. Points and triangles indicated mTBI individuals in the acute phase and HCs respectively, and lines were regression lines for each group (mTBI = red; healthy control = green), with 95% con dence interval displayed. Both two regression lines of chronological age and brain-predicted age were tted well (HCs: R2 = 0.94, P < 0.001; acute mTBI: R2 = 0.87, P < 0.001). The solid black line (black) was the line of identity (y = x). (B) The cross-sectional comparison (with error bar) of brain-PAD scores between acute mTBI patients and HCs. The brain-PAD scores in acute mTBI individuals (2.59 ± 5.97 years) were signi cantly higher than both zero (P < 0.001) and the brain-PAD scores in HCs (0.12 ± 3.19 years, P = 0.008). (C) The longitudinal comparison of brain-PAD scores between the acute and chronic phases in patients with mTBI. The brain-PAD scores in mTBI patients were still signi cantly higher than zero in their chronic phase (2.50 ± 4.54 years, P < 0.001), but not signi cantly higher than scores in the acute phase (1.87 ± 5.66 years, P = 0.51). Abbreviations: brain-PAD = brain-predicted age difference; HC = healthy control; mTBI = mild traumatic brain injury.