Neuroanatomical Substrates and Symptoms Associated With Magnetic Resonance Imaging of Patients With Mild Traumatic Brain Injury

Key Points Question What neuroanatomical changes are associated with symptoms after mild traumatic brain injury (mTBI), and when is the optimal time for acute imaging? Findings In this multicenter cohort study, 81 patients with mTBI underwent advanced magnetic resonance imaging within 72 hours and 2 to 3 weeks after injury. White matter volume and integrity evolved during that window in tandem with symptoms and were most closely associated with clinical recovery if imaging was performed within 72 hours. Meaning These findings suggest that white matter injury is associated with symptoms after mTBI and could, if detected early, help select patients at risk of poor outcome for clinical follow-up or interventional trials.

We performed a compositional data analysis as described by Aitchinson. In short, first the within-patient change of each ROI relative to the patient's intracranial volume ("perturbation") is calculated. To account for the correlation between ROIs, all perturbations for each patient were transformed using an additive log ratio (alr) resulting in one alr vector per patient. To decide whether there is a compositional change between MR1 and MR2, we tested whether the mean of all alr vectors differed from zero using the Hotelling's one-sample T 2 test for multivariate data. Note that since serial scans of the same patient were always performed on the same scanner, there were no scanner differences to adjust for.

Within patients, did
individual ROIs change in volume between MR1 and MR2 (or MR2 and MR3)?
univariate All 15 ROIs (grey and white matter) For each ROI, the within-patient change was summarised in a single value as log(Volume on MR2/Volume on MR1) negating the need for a two-sample or paired t-test. The onesample t-test assessed whether the mean change of all patients differed significantly from zero. The transformation into a log-ratio also ensured that data was normally distributed. Note that since serial scans of the same patient were always performed on the same scanner, there were no scanner differences to adjust for. 4. Did ROI volumes differ between patients and controls at MR1 (or at MR2 or at MR3)?
univariate Those 3 ROIs that changed between MR1 and MR2 mixed model ROI volume was first normalised for each person's total intracranial volume (by taking the ratio ROI/total intracranial volume) and then modelled as follows: Log(Volume) ~ group + age + sex + (1|scanner), where "group" categorised each person as either patient or control. We tested if "group" was significant according to p-values generated via Satterthwaite's degrees of freedom method (package lmerTest 3.1-2). Note that "scanner" refers to individual machines, not just scanner models, so that there are no residual site effects even if two sites used the same model. Assumptions were tested using diagnostic plots. The y-variable was binary: favorable recovery (GOSE = 8) vs. unfavorable recovery (GOSE < 8). The x-variables included age (continuous), sex (binary) and, where appropriate,"lesion presence" obtained from structured radiology reports, "WM volume" obtained from T1 imaging, and "fa tracts", "md tracts" and "both tracts" from DTI imaging. "Lesion presence" is a binary variable indicating the presence or absence of any visible lesion on any available sequence. "WM volume" is a continuous variable describing by how many standard deviations the patient's cerebral WM volume (normalised for their total brain volume) deviated from the mean of controls scanned on the same machine. The DTI variables where nominal and counted how many of the 72 tracts in each patient where abnormal with respect to only FA, only MD or both. Abnormal meant >2SD below (for FA) or above (for MD) the control mean. This binary classification resulted in better model performance than classifying FA (or MD) as high/normal/low and allowed the inclusion of the variable "both tracts" without resulting in multicollinearity as measured by the generalized variance-inflation factor corrected by the number of degrees of freedom ©2021 Richter S et al. JAMA Network Open.
12. Which imaging timepoint and sequences is more closely associated with outcome?
n/a n/a AUC, CV, AIC The above logistic regression models where compared using three measures: the area under the receiver operating characteristic curve (AUC), ten-fold cross-validation (CV) and the Akaike information criterion (AIC). To obtain the AUC, observed and predicted outcome was compared for all patients with available data. Two AUCs were compared using a paired DeLong's test. The CV accuracy is the average accuracy of ten measures obtained by randomly splitting the data into ten folds and repeatedly training the model on nine folds and testing it on the remaining fold. When comparing two models based on AIC, we considered a model to fit the data significantly better if its AIC was at least 2 units lower than that of the alternative model. 13. Are conclusions from Q10-12 robust even though patients with missing outcome data were excluded from the analyses?
n/a n/a Sensitivity analysis (best-and worst-case scenario) Sensitivity analysis for Q10: Some patients had been excluded from the complete-case analysis as they were missing ΔRPQ data. For the worst-case scenario, we assumed all 10 patients had deteriorated and imputed a ΔRPQ of +5 (the median observed in the progressive injury phenotype). For the best-case scenario, we assumed all 10 patients had improved and imputed a ΔRPQ of -4.5 (the median observed in the minimal change phenotype). An ANOVA as per Q10 was then conducted for both scenarios.
Sensitivity analysis for Q11-12: Some patients were excluded from the predictions models as they were missing GOSE data. For these patients an incomplete recovery was imputed for the worst-case scenario and a complete recovery for the best-case scenario. Logistic regression and an assessment of model performance was then conducted as per Q11 and Q12.
MR1/MR2/MR3 = Magnetic resonance scan performed within 72h/at 2-3 weeks/at 3-months after injury, ROI = Region of interest, FA = fractional anisotropy, MD = mean diffusivity, WM = white matter, SD = standard deviation. Statistical significance was determined by applying a false discovery rate threshold of 5% within each question.

eFigure 2. Analysis With and Without Patients Who Have Mass Lesions on CT
Left-hand panel: original analysis of all mild TBI patients presented as Figure 2 in the main manuscript. Right-hand panel: sensitivity analysis after exclusion of patients with visible mass lesions on their initial computed tomography scan (CT) i.e. a Marshall score of 5 or 6.