Longitudinal Prediction of Treatment Outcomes in Chronic Traumatic Brain Injury Using Artificial Intelligence and Hybrid Diffusion Imaging

Abstract

Introduction:

Chronic traumatic brain injury (cTBI) is associated with long-term cognitive, emotional, and functional impairments. It presents significant diagnostic and therapeutic challenges. Few studies have examined the effectiveness of artificial intelligence (AI) models and advanced diffusion imaging metrics to predict treatment outcomes in cTBI. This study investigated whether hybrid diffusion imaging (HYDI), a technique that employs multiple diffusion analyses, including diffusion tensor imaging (DTI) and neurite orientation dispersion and density imaging (NODDI), can identify imaging biomarkers predictive of treatment response.

Methods:

We prospectively enrolled 41 patients with cTBI who underwent HYDI scans at two timepoints. Neuropsychological outcomes were classified as favorable (clinical improvement) or unfavorable (no improvement or worsening). We assessed 6 diffusion metrics across 20 white matter regions using partial correlation analysis and 5 different AI algorithms. Outcome prediction models were trained and evaluated using k-fold cross-validation to ensure generalizability.

Results:

Results revealed significantly greater post-treatment changes in NODDI indices compared with DTI indices. In particular, regional changes in axonal density showed strong correlations with improvements in cognitive function. AI models incorporating both DTI and NODDI metrics showed promise for the prediction of the Mayo-Portland Adaptability Inventory, an assessment of postinjury adjustment in daily life and community reintegration. Overall, this study highlights the potential of HYDI-derived diffusion metrics, especially NODDI, as promising biomarkers for tracking microstructural changes and predicting therapeutic outcomes in cTBI. These findings support the use of advanced diffusion imaging and AI to inform treatment strategies and monitor recovery in patients with chronic brain injury.

Introduction

Traumatic brain injury (TBI) refers to brain dysfunction caused by external force and is associated with cognitive, emotional, and other neurological deficits. Despite advancements in neuroimaging techniques for detecting and predicting TBI-related symptoms, there has been little translation to standard clinical care.¹

White matter (WM) integrity visualization is an area of intense research interest in advanced neuroimaging of TBI. Diffusion tensor imaging (DTI), which assumes Gaussian water diffusion within a single microstructural compartment, is sensitive to WM architecture but limited in complexity. Neurite orientation dispersion and density imaging (NODDI), a multicompartment model, assesses more complex, non-Gaussian diffusion properties across intracellular, extracellular, and free-water environments. In previous work, we demonstrated that NODDI outperforms DTI in distinguishing patients with chronic TBI (cTBI) from controls and correlating with neuropsychological outcomes using artificial intelligence (AI) techniques.² While diffusion kurtosis imaging can reflect tissue complexity, it lacks the specificity of NODDI in measuring neurite density and orientation.³ However, NODDI assumes axon-dominated diffusion within each voxel and may overlook gliosis and inflammation, underscoring the need for multiple advanced diffusion metrics in identifying cTBI biomarkers.^4,5

Despite progress, prognostic imaging biomarkers for cTBI remain elusive. Recovery depends on injury characteristics and premorbid health. DTI has demonstrated the sensitivity to detect axonal injury patterns, and imaging biomarkers capable of identifying patients at risk for prolonged symptoms can inform early interventions and improve recovery outcomes.^6,7

In data-driven health care, predictive models using AI algorithms to estimate outcomes have gained prominence. Machine learning (ML), a subfield of AI, uses statistical approaches to identify relationships in data from the data itself, without explicit programing, such as the incorporation of application-specific rules. Prior research shows that ML can accurately predict survival and adverse outcomes in patients with TBI, outperforming traditional regression models.⁸ Additionally, ML algorithms have been shown, in some cases, to be stronger than traditional regression models in predicting adverse outcomes after TBI.⁹

In this study, we use a hybrid diffusion imaging sequence (HYDI), multiple diffusion metrics, and a variety of regression and ML analyses to identify biomarkers for outcome prediction (prognostic biomarkers) and the prediction of treatment outcome (predictive biomarkers) within a prospective longitudinal cohort.

In our previous work, we showed that DTI and NODDI were associated with cross-sectional cognitive outcomes using a partial correlation analysis controlling for age and gender.² We have also explored the prognostic ability of these metrics to predict outcome using several ML classification techniques.¹⁰ In this work, we build upon our baseline studies through a prospective longitudinal study design. Using a HYDI sequence to acquire both DTI and NODDI in patients with cTBI undergoing treatment, we (1) investigate the evolution of WM changes in patients with chronic TBI longitudinally using DTI and NODDI; (2) explore the prognostic significance of these microstructural changes, including the ability to track treatment response; and (3) apply ML methods to these advanced diffusion biomarkers to predict longitudinal outcome. By leveraging ML models and longitudinal imaging and outcome data, we hope to assist clinicians in identifying individuals at risk for negative outcomes.

Methods

Subject cohort

All subjects gave written informed consent in person, approved by the Thomas Jefferson institutional review board. A total of 41 patients, including 18 males and 23 females (ages 33–67), who were experiencing chronic symptoms of mild TBI (mTBI), were included in this study. mTBI was defined by the Mayo Classification System for Traumatic Brain Injury, in which an injury was classified as mild if it met the following criteria: loss of consciousness <30 min, amnesia for <24 h, and no abnormal MRI findings.¹¹ Participants had to report a history of one or more prior TBI events, in addition to chronic symptoms that met the ICD-10 criteria for mTBI and lasted for at least 6 months from the most recent TBI. Baseline assessments occurred anywhere from 3 months to several years after injury (mean: 7.24 ± 7.08 years). MRI and neuropsychological assessments were performed on patients at both the baseline visit (Fig. 1A) and follow-up at 3 months after the initiation of treatment (or 3 months after the baseline visit in the case of controls) (Fig. 1C).

Fig. 1.

Longitudinal study design for treatment, neuropsychological testing, and follow-up scan. Baseline scans and neuropsychological testing (A) were performed prior to treatment. NAC, nutritional counseling, and a waitlist control were administered after tests were performed (B). After approximately 3 months, a follow-up scan and an additional round of neuropsychological testing (C) were performed. NAC, N-acetylcysteine.

Treatment and image acquisition

Patients were randomly assigned to one of three treatment groups: (1) N-acetylcysteine (NAC), a treatment aimed at boosting glutathione (GSH) levels; (2) nutritional counseling; and (3) a waitlist control group (Fig. 1B). Seventeen patients received NAC (8 M, 9 F, mean age: 55.3 ± 8.4), 15 received nutritional counseling (10 M, 5 F, mean age: 49.7 ± 14.6), and 9 control patients received no treatment (0 M, 9 F, mean age: 50.6 ± 17.3). Patients receiving NAC were treated with a combination approach over 3 months, with weekly intravenous infusions of 50 mg/kg given over approximately 1 h, and then 500 mg given orally twice per day on the other 6 days of the week. Participants in the nutritional counseling group were given an individualized 1-h counseling session focusing on nutrient-dense diets, cognitive health, and brain injury recovery. A diet program was developed and tailored to individual preferences and needs. Informational pamphlets were given, and then the patient was allowed to contact the counselor at any time over the 3-month implementation period. Check-ins at midpoint and after 3 months of treatment were performed to assess adherence to the diet. NAC acts through the upregulation of the level of GSH, a combination of l-glutamic acid, l-cysteine, and glycine, within the brain. Administration of NAC is believed to maintain high levels of GSH in the brain, which acts as a free radical scavenger and as an antioxidant.¹²

Neuropsychological tests

A battery of neuropsychological tests was performed in patients at the time of the baseline scan and at the post-treatment visit. These included (1) Trail Making Test Part A (TMT-A), which assesses cognitive processing speed by having participants connect numbers in a sequential order; (2) forward digit span (FDS) and backward digit span of the Wechsler Adult Intelligence Scale (WAIS-III) to assess working memory; and (3) self-reported symptomatology, including the Rivermead Post-Concussion Symptoms Questionnaire (RPQ-3, RPQ-13), Beck Depression Inventory (BDI), Insomnia Severity Index, and others. The Mayo-Portland Adaptability Inventory was used to assess the subjects’ range of physical, cognitive, emotional, behavioral, and social problems. This includes its three subscales: ability index, adjustment index, and participation index, which are highly developed and well documented for measuring psychometric properties. The TMT was used to assess cognitive flexibility, while the WAIS-III Digit Span Subtest was used to measure working memory.

Image acquisition

The image acquisition protocol was similar to one that was previously reported²; however, images were acquired at two timepoints: baseline and post-treatment. The HYDI sequence was performed using a 3T Siemens Biograph MR PET-MR scanner with a 32-channel head coil. For segmentation and registration of WM atlas structures, an anatomical T1-weighted image was obtained for all subjects. MRI parameters for the T1-weighted anatomical sequence were: repetition time = 1.6 s, echo time = 2.46 ms, field of view = 250 × 250 mm, matrix = 512 × 512, voxel size = 0.49 × 0.49 mm,⁵ 176 slices with slice thickness = 1 mm. To capture both high-resolution diffusion metrics and the full complexity of WM microstructure, we employed the simultaneous multi-slice (SMS) HYDI pulse sequence. The minimum b-value was 0 s/mm^2,⁵ with five concentric diffusion-weighting shells (b-values = 250, 1000, 2000, 3250, 4000 s/mm²). A total of 144 diffusion-weighting gradient directions (6, 21, 24, 30, and 61 in each shell) were encoded. MRI parameters for the HYDI sequence were as follows: repetition time = 3.17 s, echo time = 120 ms, field of view = 240 × 240 mm, matrix = 96 × 96, voxel size = 2.5 × 2.5 mm², 63 slices with slice thickness = 2.5 mm, SMS factor = 2, and total scan time of 8 min.

Diffusion processing

The diffusion MRI data were verified to be free of major artifacts or excessive movement, defined as more than 2 mm of translation and/or rotation. DTI preprocessing and analysis were performed using tools from the Oxford Centre for Functional MRI of the Brain (FMRIB) software library (FSL). Images were corrected for eddy distortions and motion using an average of the b = 0 s/mm² volumes as a reference. Registered images were skull-stripped using the Brain Extraction Tool. Brain masks were visually inspected for anatomical fidelity, and denoising was applied using MRtrix.¹³ DTI parameter maps were calculated using the FSL Diffusion Toolbox.

Multicompartment biophysical modeling of diffusion MR imaging was performed using the NODDI toolbox v0.9 (www.nitrc.org/projects/noddi_toolbox). The NODDI toolbox provides direct, specific measurements of tissue microstructure using model-based signals from Diffusion Weighted Imaging (DWI). After processing DTI and NODDI data, fractional anisotropy (FA), axial diffusivity (AD), mean diffusivity (MD), radial diffusivity (RD), volume of the intraneurite compartment (Vic), and orientation dispersion index (ODI) maps were generated. All maps were aligned into the FMRIB58 FA template in MNI152 standard space using the nonlinear registration algorithm FMRIB's Non-linear Image Registration Tool (FNIRT).

WM regions-of-interest-based analysis

Twenty main WM tracts were studied using masks from the Johns Hopkins University (JHU) WM tractography atlas, mapped onto MNI152 space, and resampled to 1-mm resolution. Binary masks from each region of interest (ROI) were used to mask the normalized maps previously registered to standard space using nonlinear tools in the tract-based spatial statistics procedure. Mean FA, AD, MD, RD, Vic, and ODI were obtained from each subject’s normalized maps for each ROI at baseline and post-treatment. Changes in diffusion metrics were calculated by subtracting baseline values from post-treatment values for each region.

To investigate relationships between neuropsychological improvements and diffusion metrics, partial correlations controlling for age and gender were computed. This helped to isolate diffusion effects and minimize confounders. A Spearman rank correlation was performed using in-house MATLAB code, calculating partial correlation coefficients between neuropsychological improvement and regional diffusion metrics (baseline, post-treatment, and change), controlling for age and gender.

The Spearman rank correlation was used to assess the strength and direction of monotonic relationships between neuropsychological improvements and diffusion metrics (Fig. 2).

Fig. 2.

Spearman rank correlation was performed on baseline and post-treatment scans to define predictive biomarkers (A) and correlation between changes in diffusion metrics and neuropsychological outcomes (B).

ML analysis

Five classical ML models were adopted using classification algorithms, including support vector machine (SVM), K-nearest neighbors (KNN), logistic regression (LR), random forest (RF), and decision tree (DT). SVM searches for an optimal hyperplane between classes, maximizing the classification margin. LR predicts relationships between dependent and independent variables. KNN predicts class by calculating the distance between test data and training points. RF creates accurate predictions by combining multiple weak rules. All ML algorithms were implemented using Scikit-Learn in Python (v0.22.1). Models were trained and tested using T1, DTI (FA, AD, MD, RD), and NODDI (ODI, Vic) metrics, evaluated using accuracy, ROC curves, and confusion matrices.

The use of classification models for predicting patient response to treatment by simplifying treatment response as either “favorable” or “unfavorable” allowed the authors to leverage longitudinal data to predict disease response to differing treatment types. Outcomes were classified as (1) favorable: improved neuropsychological outcome in response to treatment, and (2) unfavorable: worsening or no change in neuropsychological outcome. The binary classification criteria were used as inputs for ML algorithms to train and test the models. This enabled the creation of a dataset for classifier training. Features included DTI and NODDI values from baseline and post-treatment scans within 20 JHU regions. Changes in DTI and NODDI values were calculated by subtracting baseline from post-treatment values (Fig. 3). Six classification models were developed using DTI and NODDI parameters to predict baseline, post-treatment, and improvement in neuropsychological outcomes.

Fig. 3.

Machine learning-based analysis of diffusion parameters. Six classification algorithms (LR, DT, RF, KNN, SVM) were trained using k-fold cross-validation and tested. Features included DTI and NODDI values from the baseline scan, post-treatment scan, and changes between (post-treatment values subtracted by the baseline values). For example, a patient’s pre-treatment FA, post-treatment FA, then the post-treatment FA value subtracted by the pre-treatment FA value. The target variable to predict was whether the patient had a favorable versus unfavorable neuropsychological outcome prior to receiving treatment. DT, decision tree; DTI, diffusion tensor imaging; FA, fractional anisotropy; KNN, K-nearest neighbors; LR, logistic regression; NODDI, neurite orientation dispersion and density imaging; RF, random forest; SVM, support vector machine.

Results and Discussion

Patient outcomes and response to treatment types

Overall, the cohort demonstrated improvements across most neuropsychological assessments, though some individual patients showed decreased performance in certain areas, such as FDS and vigor. On average, patients showed a 2.9–5.9% reduction in the time it took to make trails. Moreover, patients exhibited improvements across all neuropsychological assessments, particularly in areas related to participation and adjustment, as measured by the Mayo-Portland Adaptability Index. Notably, the only areas showing decreased performance were measures of vigor and the FDS.

Changes in diffusion metrics

On average, at post-treatment, DTI metrics showed slight increases across all regions, while NODDI metrics demonstrated significantly higher percent changes. Notably, Vic exhibited the largest increase (2.022–12.336%), while ODI showed a considerable decrease (7.254–16.424%) after treatment (Fig. 4).

Fig. 4.

Percent change in all six diffusion metrics from baseline to post-treatment, averaged across all regions and all patients.

Longitudinal correlation to predict symptom improvement

In our analysis, the RPQ-13 questionnaire was used to assess post-concussion symptoms, with total scores for symptoms such as headache, fatigue, and vigor. The BDI was used to assess depressive symptoms, and state anxiety was measured using the State–Trait Anxiety Inventory.

To evaluate whether diffusion metrics correlated with neuropsychological improvements over time, we report (Table 1) only correlations significant at both baseline and post-treatment. After False Discovery Rate (FDR) correction, FA and RD demonstrated significant correlations with at least one outcome measure (Fig. 5). FA positively correlated with depression, Mayo-Portland Adaptability Inventory, and headache symptom improvement, while RD showed negative correlations with adjustment and ability. Vic (axonal density) consistently correlated with Mayo-Portland and RPQ-13 scores over time (Fig. 6). FA in the inferior fronto-occipital fasciculus, superior longitudinal fasciculus, and uncinate fasciculus demonstrated longitudinal correlations. Baseline RD in the inferior longitudinal fasciculus and forceps minor also demonstrated longitudinal correlations. Baseline Vic showed strong correlations in the inferior longitudinal fasciculus and temporal regions of the superior longitudinal fasciculus.

Fig. 5.

Correlations that were significant (FDR-corrected) for both the baseline (B) and post-treatment (P) time points between neuropsychological improvement and DTI metrics.

Fig. 6.

Correlations that were significant (FDR-corrected) for both the baseline and post-treatment time points between neuropsychological improvement and metrics of the NODDI model.

Table 1.

Regional Midpoint and Baseline DTI and NODDI Metrics Significantly Correlated with Improvements in Neuropsychological Outcomes

Region	Neuropsychological outcome	r-Baseline	p-Value(corrected)	r-Post	p-Value (corrected)
Fractional anisotropy
Inferior fronto-occipital fasciculus R	Beck depression	0.405	0.020 (0.027)	0.352	0.028 (0.077)
Inferior fronto-occipital fasciculus R	Mayo-Portland	0.368	0.021 (0.029)	0.411	0.009 (0.044)
Superior longitudinal fasciculus L	Mayo-Portland	0.385	0.015 (0.029)	0.437	0.005 (0.047)
Uncinate fasciculus L	Beck depression	0.368	0.021 (0.030)	0.430	0.006 (0.047)
Superior longitudinal fasciculus (temporal part) L	Headache	0.395	0.013 (0.026)	0.414	0.009 (0.046)
Radial diffusivity
Inferior longitudinal fasciculus R	Adjustment index	−0.395	0.013 (0.055)*	−0.427	0.007 (0.028)
Forceps minor	Ability index	−0.475	0.002 (0.031)	0.441	0.005 (0.026)
Forceps minor	Adjustment index	−0.449	0.004 (0.046)	−0.387	0.015 (0.028)
Axonal density (Vic)
Inferior longitudinal fasciculus L	Mayo-Portland	0.500	0.001 (0.015)	0.344	0.032 (0.056)*
Inferior longitudinal fasciculus R	Mayo-Portland	0.538	>0.001 (0.018)	0.378	0.018 (0.067)*
Superior longitudinal fasciculus L	Mayo-Portland	0.501	0.001 (0.017)	0.351	0.028 (0.054)*
Superior longitudinal fasciculus (temporal part) L	Mayo-Portland	0.449	0.004 (0.030)	0.364	0.023 (0.062)*
Superior longitudinal fasciculus (temporal part) L	RPQ-13	0.434	0.006 (0.035)	0.420	0.008 (0.074)*
Superior longitudinal fasciculus (temporal part) R	Mayo-Portland	0.492	0.001 (0.016)	0.437	0.005 (0.102)
Superior longitudinal fasciculus (temporal part) R	RPQ-13	0.457	0.003 (0.030)	0.408	0.010 (0.063)*

Correlation values and p-values are reported for both baseline and midpoint scans (*corrected p-value is nearly significant).

DTI, diffusion tensor imaging; NODDI, neurite orientation dispersion and density imaging; RPQ, Rivermead Post-Concussion Symptoms Questionnaire.

Changes in diffusion metrics and recovery

Of the six diffusion metrics, only the change in Vic from baseline to post-treatment scans in certain regions showed significant correlations with symptom improvements (Table 2 and Fig. 7). Associations with improvements in headache were found in the uncinate fasciculus and superior longitudinal fasciculus. Associations with improvements in the Mayo-Portland Adaptability Index were found in regions of the uncinate fasciculus, superior longitudinal fasciculus, inferior fronto-occipital fasciculus, and inferior longitudinal fasciculus. Last, improvement of the RPQ-13 was associated with Vic in the inferior longitudinal fasciculus, inferior fronto-occipital fasciculus, and uncinate fasciculus.

Fig. 7.

Regional changes in axonal density (Vic) that are significantly associated with improvements in headache (A), Mayo-Portland (B), and RPQ-13 (C). Significant correlations between axonal density (Vic) and symptoms were found in the following regions: superior longitudinal fasciculus (SLF), uncinate fasciculus (UNC), inferior longitudinal fasciculus (ILF), and inferior fronto-occipital fasciculus (IFO).

Table 2.

Table Showing Significant Correlations between Change in Regional Axonal Density (Vic) and Neuropsychological Outcomes, after False Discovery Rate Correction

Region	Neuropsychological outcome	Correlation	p-Value (corrected)
Inferior fronto-occipital fasciculus R	Mayo-Portland	−0.443	0.005 (0.040)
Inferior fronto-occipital fasciculus R	Mayo-Portland	−0.401	0.011 (0.043)
Superior longitudinal fasciculus R	Mayo-Portland	−0.386	0.015 (0.047)
Inferior longitudinal fasciculus R	RPQ-13	−0.479	0.002 (0.023)
Uncinate fasciculus R	Mayo-Portland	−0.436	0.005 (0.031)
Superior longitudinal fasciculus R	RPQ-13	−0.482	0.002 (0.031)
Superior longitudinal fasciculus R	Headache	−0.436	0.005 (0.031)
Uncinate fasciculus R	RPQ-13	−0.534	<0.001 (0.015)
Uncinate fasciculus L	Headache	−0.413	0.009 (0.043)
Uncinate fasciculus R	Headache	−0.395	0.013 (0.043)
Superior longitudinal fasciculus (temporal part) L	Forward	0.412	0.009 (0.039)

ML analysis

We employed five different classification methods to predict patient outcomes after treatment for cTBI, using baseline, post-treatment, and change metrics from DTI and NODDI as features. The five models (LR, DT, RF, KNN, and SVM) were compared based on mean accuracy (Fig. 8).

Fig. 8.

Machine learning results of baseline (B), midpoint (M), and change (□) in diffusion values using all 20 regions of the JHU atlas to predict neuropsychological improvements. Algorithms include logistic regression (LR), decision tree (DT), random forest (RF), K-nearest neighbors (KNN), and support vector machine (SVM).

The best-performing model, LR, achieved an average accuracy, over all outcome measures and all three measures (baseline, post-treatment, and change from baseline to post-treatment) of 56.1%, while DT models demonstrated the lowest average accuracy at 51.9%. However, in this exploratory analysis, prediction of adjustment and participation based on the Mayo-Portland Adaptability Inventory was better than prediction of other outcomes. The results, while preliminary, show promise for using ML to predict neuropsychological recovery in chronic TBI.

We found widespread regions with significant changes in conventional and higher-order DTI and NODDI parameters of patients with symptoms of chronic TBI. Among the baseline, post-treatment, and post-baseline values of all diffusion metrics, the post-baseline values were the most promising. We found that FA provided up to 76% accuracy in predicting longitudinal improvements in depression. Interestingly, FA was highly correlated in specific regions with the BDI based on the Spearman rank correlation.

Discussion

This study provides a novel, prospective, longitudinal analysis of changes in HYDI parameters in patients with cTBI. From our analysis, we drew three main conclusions: 1.

NODDI metrics, particularly Vic, show promise as prognostic markers of patient outcomes, potentially outperforming DTI.

Vic was also a promising biomarker for tracking recovery, with regional changes in Vic from the baseline to post-treatment scans correlating strongly with symptom improvement.

ML classification models based on both DTI and NODDI metrics predicted improvement of the Mayo-Portland Adaptability Index better than improvements in other outcome measures. This result, while preliminary, could be validated in a larger and/or independent sample.

For many mild TBI injuries, pathological changes postinjury may be transient or can persist. Cellular injury during the acute TBI phase is volatile and highly dynamic.¹⁴ Current studies typically focus on the acute phase, where bidirectional changes are often undetectable. For the first time, we analyzed DTI and NODDI biomarkers to detect longitudinal symptom improvements in patients treated for cTBI. Although no significant differences in outcomes were detected across the three treatment groups, the techniques and analysis described here can be applied to other therapeutic trials with larger sample sizes.

We observed increasing Vic and decreasing ODI values over time, suggesting progressive axonal degeneration and contrasting with minimal changes in DTI metrics. These results suggest that NODDI, particularly Vic and ODI, are more sensitive to WM axonal loss than DTI, with treatment effects possibly tied to changes in neurite density. This supports prior studies suggesting NODDI is a more specific biomarker than FA.

Using Spearman rank partial correlation, we found that Vic values were significantly correlated with a patient’s improvement on neuropsychological measures over time, increasing or decreasing in parallel with symptom changes. This supports prior research in which FA values remained persistently reduced in certain tracts.^15,16 Both FA and Vic reflect aspects of axonal density, though FA has long been debated in terms of its main neurobiological contributor, with two primary candidates: tract density and axonal arrangement.^17–24 Our findings suggest that, in the context of TBI, axonal density (as measured by Vic) may be superior for longitudinal monitoring of symptoms over time than FA. Although FA reflects tract integrity and axonal arrangement, its role as a reliable biomarker has been debated. In contrast, Vic more directly measures axonal density, which may make it a more sensitive biomarker for detecting neurobiological changes and tracking recovery.

Several limitations of this study must be considered. One caveat is that changes in NODDI metrics can be, in part, due to issues around the test–retest reliability of these measures, reported by others.^16,25,26 The absence of a healthy control group limits our ability to directly compare DTI and NODDI metrics across individuals without a history of TBI. Additionally, treatment groups in the patient cohort were unevenly distributed, which may have contributed to variability in the performance of the ML models. Future directions include the study of a larger and more balanced population, conduction of treatment-specific analyses, and test–retest reliability evaluation of the NODDI measures. Finally, due to the limited sample size of longitudinal HYDI data, we could not include more complex ML classifiers like multilayer perceptron or adaptive boosting. Future work will involve larger datasets and the exploration of alternative classifiers.

Conclusion

Advanced diffusion MRI techniques, including DTI and NODDI, provide valuable insights into brain microstructure following cTBI. While NODDI proved more sensitive to detecting changes in symptoms over time, both DTI and NODDI hold significant promise for integration into scientific models that could assist clinicians in identifying patients at risk of chronic symptoms and predicting their response to treatment. Moving forward, translating these findings into clinical applications is critical, particularly as current diagnostic and detection methods for TBI are limited.

Footnotes

Authors’ Contributions

J.J.M.: Conceptualization, data curation, formal analysis, investigation, methodology, software, validation, visualization, writing—original draft, writing—review and editing, and supervision. R.W.: Conceptualization, data curation, formal analysis, investigation, methodology, resources, software, validation, visualization, writing—original draft, and writing—review and editing. D.M.: Conceptualization, data curation, formal analysis, methodology, resources, supervision, and validation. M.A.: Conceptualization, data curation, methodology, resources, software, supervision, and writing—review and editing. K.C.K.: Data curation, formal analysis, investigation, software, validation, writing—original draft, and writing—review and editing. R.H.: Data curation, formal analysis, investigation, methodology, visualization, writing—original draft, and writing—review and editing. G.Z.: Conceptualization, data curation, project administration, resources, software, validation, and visualization. C.H.: Data curation, investigation, methodology, validation, visualization, and writing—review and editing. E.N.: Data curation, formal analysis, investigation, project administration, and writing—review and editing. N.W.: Conceptualization, data curation, project administration, resources, and writing—review and editing. A.J.B.: Conceptualization, data curation, project administration, resources, supervision, and writing—review and editing. C.W.: Data curation, investigation, project administration, resources, supervision, and writing—review and editing. D.A.M.: Conceptualization, funding acquisition, project administration, resources, supervision, and writing—review and editing. X.J.: Conceptualization, data curation, investigation, project administration, resources, software, supervision, validation, writing—original draft, and writing—review and editing. Q.W.: Conceptualization, data curation, funding acquisition, investigation, project administration, resources, supervision, writing—original draft, and writing—review and editing. A.B.N.: Conceptualization, data curation, funding acquisition, methodology, project administration, resources, supervision, writing—original draft, and writing—review and editing. F.B.M.: Conceptualization, data curation, formal analysis, funding acquisition, investigation, methodology, project administration, resources, software, supervision, validation, writing—original draft, and writing—review and editing.

Author Disclosure Statement

No competing financial interests exist.

Funding Information

Funding was provided by the Marcus Foundation.

References

Carvalho

, Chowdhury

, Zaninotto

, et al. Case report: Improving verbal retrieval deficits with high definition transcranial direct current stimulation targeting the pre-supplementary motor area in a patient with chronic traumatic brain injury. Front Neurol, 2021; 1:678518; doi: 10.3389/fneur.2021.678518

Muller

, Middleton

, Alizadeh

, et al. Hybrid diffusion imaging reveals altered white matter tract integrity and associations with symptoms and cognitive dysfunction in chronic traumatic brain injury. Neuroimage Clin, 2021; 30(April):102681; doi: 10.1016/j.nicl.2021.102681

Fieremans

, Jensen

, Helpern

. White matter characterization with diffusional kurtosis imaging. Neuroimage, 2011; 58(1):177–188; doi: 10.1016/j.neuroimage.2011.06.006

Kinnunen

, Greenwood

, Powell

, et al. White matter damage and cognitive impairment after traumatic brain injury. Brain, 2011; 134(Pt 2):449–463; doi: 10.1093/brain/awq347

Kraus

, Susmaras

, Caughlin

, et al. White matter integrity and cognition in chronic traumatic brain injury: A diffusion tensor imaging study. Brain, 2007; 130(Pt 10):2508–2519; doi: 10.1093/brain/awm216

Ling

, Peña

, Yeo

, et al. Biomarkers of increased diffusion anisotropy in semi-acute mild traumatic brain injury: A longitudinal perspective. Brain, 2012; 135(Pt 4):1281–1292; doi: 10.1093/brain/aws073

Kontos

, Kotwal

, Elbin

, et al. Residual effects of combat-related mild traumatic brain injury. J Neurotrauma, 2013; 30(8):680–686; doi: 10.1089/NEU.2012.2506

Khalili

, Rismani

, Nematollahi

, et al. Prognosis prediction in traumatic brain injury patients using machine learning algorithms. Sci Rep, 2023; 13(1):960; doi: 10.1038/s41598-023-28188-w

Courville

, Kazim

, Vellek

, et al. Machine learning algorithms for predicting outcomes of traumatic brain injury: A systematic review and meta-analysis. Surg Neurol Int, 2023; 14:262; doi: 10.25259/SNI_312_2023

10.

Muller

, Wang

, Milddleton

, et al. Machine learning-based classification of chronic traumatic brain injury using hybrid diffusion imaging. Front Neurosci, 2023; 17:1182509; doi: 10.3389/fnins.2023.1182509

11.

Malec

, Brown

, Leibson

, et al. The mayo classification system for traumatic brain injury severity. J Neurotrauma, 2007; 24(9):1417–1424; doi: 10.1089/neu.2006.0245

12.

Eakin

, Baratz-Goldstein

, Pick

, et al. Efficacy of N-Acetyl Cysteine in traumatic brain injury. PLoS One, 2014; 9(4):e90617; doi: 10.1371/journal.pone.0090617

13.

Tournier

, Calamante

, Connelly

. MRtrix: Diffusion tractography in crossing fiber regions. Int J Imaging Syst Tech, 2012; 22(1):53–66; doi: 10.1002/ima.22005

14.

Yin

, Li

D-D

, Huang

, et al. Longitudinal changes in diffusion tensor imaging following mild traumatic brain injury and correlation with outcome. Front Neural Circuits, 2019; 13:28; doi: 10.3389/fncir.2019.00028

15.

Maruta

, Palacios

, Zimmerman

, et al. Chronic post-concussion neurocognitive deficits. i. relationship with white matter integrity. Front Hum Neurosci, 2016; 10(FEB2016):35–38; doi: 10.3389/fnhum.2016.00035

16.

Palacios

, Owen

, Yuh

, et al.; TRACK-TBI Investigators. The evolution of white matter microstructural changes after mild traumatic brain injury: A longitudinal DTI and NODDI Study. Sci Adv, 2020; 6(32):eaaz6892; doi: 10.1126/sciadv.aaz6892

17.

Friedrich

, Fraenz

, Schlüter

, et al. The relationship between axon density, myelination, and fractional anisotropy in the human corpus callosum. Cereb Cortex, 2020; 30(4):2042–2056; doi: 10.1093/cercor/bhz221

18.

Refaeilzadeh

, Tang

, Liu

. Cross-Validation. In: Encyclopedia of Database Systems. ( LIU

, ÖZSU

., eds.) Springer: Boston, MA; 2009; doi: 10.1007/978-0-387-39940-9_565

19.

Anguita

, Ghelardoni

, Ghio

, et al. The ‘K’ in K-fold Cross Validation. Balint

, Antala

, Carty

, Mabieme

JMA

, Amar

, Kaplanova

, eds. Uniwersytet śląski; 2022:441–446. doi:10.2/JQUERY.MIN.JS

20.

Myles

, Feudale

, Liu

, et al. An introduction to decision tree modeling. J Chemom, 2004; 18(6):275–285; doi: 10.1002/CEM.873

21.

Rodriguez-Galiano

, Ghimire

, Rogan

, et al. An assessment of the effectiveness of a random forest classifier for land-cover classification. ISPRS J Photogrammetry Remote Sensing, 2012; 67(1):93–104; doi: 10.1016/J.ISPRSJPRS.2011.11.002

22.

Palczewska

, Palczewski

, Robinson

, et al. Interpreting random forest classification models using a feature contribution method. Advances in Intelligent Systems and Computing, 2014; 263:193–218; doi:10.1007/978-3-319-04717-1_9

23.

Suthaharan

. Support vector machine. In: Machine Learning Models and Algorithms for Big Data Classification. Integrated Series in Information Systems, vol 36. Springer: Boston, MA; 2016. 207–235. doi:10.1007/978-1-4899-7641-3_9

24.

Zhang

, Du

, Wang

. Composite Quantization for Approximate Nearest Neighbor Search. 2014. 838–846. Available from: https://proceedings.mlr.press/v32/zhangd14.html [Last accessed: August 31, 2022].

25.

Andica

, Kamagata

, Hayashi

, et al. Scan–rescan and inter-vendor reproducibility of neurite orientation dispersion and density imaging metrics. Neuroradiology, 2020; 62(4):483–494; doi: 10.1007/s00234-019-02350-6

26.

Kamiya

, Hori

, Aoki

. NODDI in clinical research. J Neurosci Methods, 2020; 346:108908; doi: 10.1016/j.jneumeth.2020.108908