Traumatic brain injury (TBI) is a common condition with many potential acute and chronic neurological consequences (Smith et al., 2019), contributing to ~1 million deaths in the United States over the last two decades (Daugherty and Zhou, 2016). The neuropathology of chronic traumatic brain injury (cTBI) consists of a primary injury that is a direct consequence of traumatic insult, and a secondary injury that results from a cascade of molecular and cellular events including cell death, axonal injury, and inflammation (Anguita et al., 2022; Zhang et al., 2022). To better understand the underlying neuropathological mechanisms, there remains a growing need for information on the chronic effects of TBI (Wickwire et al., 2016).

Neuroimaging plays a critical role in the acute setting of brain injury, both in diagnosis and in guiding appropriate management by detecting injuries that require intervention or monitoring (Taylor and Gercel-Taylor, 2014; Douglas et al., 2015; Mckee and Daneshvar, 2015). However, in the setting of most mild-to-moderate injury, conventional T1-weighted imaging is typically normal (McCrory et al., 2009). Moreover, an initial assessment of TBI severity does not necessarily predict the extent of chronic disability (National Academies of Sciences, 2019). Therefore, advanced neuroimaging biomarkers are being actively researched in an attempt to better diagnose and monitor the acute and chronic effects of TBI (Hu et al., 2022).

Diffuse axonal injury is thought to be a key pathological mechanism underlying TBI and, as a result, it has led toward the development of advanced MR techniques for the visualization of WM integrity (Hashim et al., 2017). DTI and neurite orientation dispersion imaging (NODDI) (Zhang et al., 2012) are advanced MR techniques, which are thought to reflect the integrity of microstructural properties of the white matter (WM) in a range of clinical conditions. Diffusion tensor imaging (DTI) assumes Gaussian diffusion within a single microstructural compartment, whereas NODDI probes more complex, non-Gaussian properties using high-performance magnetic field gradients (Kamiya et al., 2020). Unlike DTI, NODDI uses seven parameters to measure the properties of the WM microstructure including intracellular, extracellular, and free water, whereas DTI is limited in its description of isotropic vs. anisotropic diffusion for a particular voxel (Muller et al., 2021).

Previously it has been shown that DTI and NODDI offer different, yet complementary, information regarding the microstructural integrity in patients with acute to chronic TBI (Wu et al., 2018; Palacios et al., 2020; Muller et al., 2021). Among DTI metrics, fractional anisotropy (FA) is the most studied and is often used as an indicator of white matter “integrity.” FA is mathematically defined as the square root of the sum of the squares of the eigenvalues within a diffusion tensor. Therefore, it is driven by axial diffusivity (AD; principle eigenvalue) and the second and third eigenvalues known as radial diffusivity (RD). DTI has been shown to be sensitive to acute WM changes, whereas NODDI has been shown to be more sensitive in detecting axonal degeneration over time (Timmers et al., 2016). Although DTI and NODDI may be more sensitive and specific for diagnosing TBI, they are rarely acquired in clinical practice, as advanced imaging techniques require additional scanning time, posing difficult challenges for practical application. To overcome this, a hybrid diffusion imaging (HYDI) sequence has been developed which acquires data that enable complementary post–data-processing strategies including DTI and NODDI (Wu and Alexander, 2007).

Machine learning (ML) methods have been demonstrated to be effective for various medical purposes (Chong et al., 2015; Vergara et al., 2017; Mohamed et al., 2022). The results of prior studies have suggested that biomarkers obtained from diffusion data, such as NODDI and DTI, combined with machine learning classification have the potential to be used for detecting neurodegenerative diseases such as Alzheimer's (Prasuhn et al., 2020) and Parkinson's (Pease et al., 2022) abnormalities. However, to the best of our knowledge, little research is conducted on using machine learning for inferring cognitive deficits in patients who experience cTBI (Qu et al., 2021). Currently, available algorithms lack the capabilities to make predictions in the absence of age-matched controls (Minaee et al., 2017) or fail to explore the diagnostic potential of higher-order diffusion models, such as NODDI (Qu et al., 2021). Notably, DTI and NODDI provide metrics with similar biological interpretations. These metrics are promising biomarkers for symptom development in TBI. Thus, the purpose of this study was to develop an ML model using various DTI and NODDI metrics fit from HYDI data, for inferring outcomes in patients with cTBI, and to compare its performance with conventional T1-weighted imaging.

This study was approved by the institutional review board of Thomas Jefferson University Hospital. All subjects included in the study provided informed consent.

As the trail-making task has been shown to be a reliable indicator of potential signs of cognitive impairment, we used the time it took for patients to complete trail making A and B (seconds) to evaluate the diagnostic potential of the machine learning algorithms. A longer time to complete trail making indicates worse performance for this task. Age, sex, number of TBI, and injury-to-imaging interval did not have a significant prediction on the classification of symptoms, indicating no significant correlation or confounding effects. The experimental results are indicated in Figures 2, 3.

To the best of our knowledge, this is the first study to describe the development of an ML-based classification algorithm to compare DTI and NODDI, in a real clinical setting by analyzing the same data, for the diagnosis of cTBI. In clinical practice, it is difficult to infer symptoms of TBI using conventional MR imaging, though several studies using DTI and NODDI have been shown to correlate with symptoms (Palacios et al., 2020). MR imaging findings in cTBI are often subtle, with DTI and NODDI rarely being acquired, and visual assessments of these findings being subjective (Zhang et al., 2012). Moreover, ML models have not yet been widely developed and verified within the field of diagnosis, creating a need for ML methods to infer symptom pathology from neuroimaging, particularly within a real-world clinical setting. We found that DTI and NODDI resulted in mean accuracies of 58.7–73.0% and 64.0–72.3%, respectively, and consistently outperformed T1-weighted imaging in all ML algorithms across multiple testing methods to infer symptoms of cognitive impairment. Our results are informative for the community and elucidate whether NODDI and DTI detect changes in similar locations, are more predictive than conventional T1-weighted imaging, and whether there are advantages to acquiring an advanced diffusion sequence for the inference of cTBI.

The DT feature-reduction method developed in this study successfully improved the performance of K-NN, with an 11% mean improvement in inference accuracy. This suggests that the effects of cTBI may be more localized to specific brain regions, rather than widespread throughout the whole brain. The results of the DT feature ranking method improving the predictive accuracy suggests that clinicians should focus on specific regions when looking for signs of clinical impairment as more specific areas of the brain appear to be correlated with clinical outcomes. Our data are generally consistent with previous findings, where importance-based variable selection has been shown to allow for simplified predictive models to be created while maintaining inference accuracy. In a diverse unpredictable condition, such as cTBI, these results help refine our approach to head injury assessment, decision-making, and outcome inference targeted at model sensitivity and specificity.

This study has several limitations. First, this study was based on a relatively small subject cohort for the development of an ML algorithm. A larger subject cohort including a wider range of outcomes from multiple centers is desirable to increase the generalizability and robustness of the algorithm. This study, however, was not intended to develop an all-inclusive tool to differentiate various causes of cognitive impairment but rather to compare the predictive ability of DTI, NODDI, and T1-weighted imaging. The HYDI data collected were done so in a timeframe that may be clinically feasible for most clinical centers. Further validation with larger, prospectively collected test datasets may be necessary to determine whether this method is applicable to various types of cognitive impairment. Second, diagnostic criteria for cTBI were based on clinical diagnosis and self-reported symptoms; this is not only a weakness but also a strength as this is more generalizable to what would be experienced, given a real-world clinical setting. However, more robust inferences could be generated by using more quantitative measurements of cognitive abilities.

Finally, our preliminary experiment selects feature engineering-based approaches and utilizes DT for feature ranking and traditional ML for testing. In future studies, we will consider more sophisticated feature ranking methods, such as Lasso regression, and more powerful classifiers including multilayer perceptron (MLP) and graph neural networks (GNNs).

The clinical applications of the development of our ML models are extensive, as it not only indicates that deficits in cognition can be predicted in cTBI patients in the absence of age-and-sex-matched healthy controls but also highlights the utility of adopting advanced diffusion imaging into a clinical pipeline of what is traditionally used. Our findings could affect not only the diagnosis and treatment monitoring of patients with cTBI but also could offer a method of determining a patient's likelihood of developing cognitive deficits within the acute stages of injury.

ML-based classification algorithms using DTI and NODDI consistently outperform conventional T1-weighted brain imaging for predicant patients' symptoms and unfavorable outcomes 6 months following the traumatic incident. Reducing the number of features improves the accuracy of inference. These results indicate that ML-based classification built on higher-order diffusion modeling and reduced features may be a promising tool for inferring cognitive deficits in TBI, improving clinical decision-making and yielding better patient outcomes.

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author.

The studies involving human participants were reviewed and approved by IRB. Written informed consent to participate in this study was provided by the participants' legal guardian/next of kin.

JM and RW: conceptualization, methodology, investigation, software, validation, and writing-original draft preparation. DMi and MA: conceptualization, data curation, and software. KK and RH: data curation, formal analysis, and investigation. GZ, NW, CW, DMo, and AB: resources, supervision, and project administration. CH and EN: resources, data curation, and project administration. XJ: conceptualization, methodology, formal analysis, investigation, software, resources, and writing - review and editing. QW: conceptualization, methodology, investigation, resources, supervision, and writing - review and editing. AN: conceptualization, methodology, investigation, software, resources, supervision, project administration, and writing - review and editing. FM: conceptualization, methodology, investigation, formal analysis, investigation, software, resources, project administration, supervision, and writing - review and editing.