Automated Assessment of White Matter Integrity in TBI Using Machine Learning

Michael E Debakey Va Medical Center, Houston TX

Investigators

Abstract

DESCRIPTION (provided by applicant): Mild traumatic brain injury (mTBI) is the signature injury of the wars in Afghanistan and Iraq. Recent statistics indicate that 60% of blast injuries result in TBI and approximately 20% of returning OEF/OIF Veterans have sustained a TBI, with the majority classified as mTBI. Although many sequelae of mTBI resolve within a few months, a substantial portion of patients experience difficulties for years. Diagnosis of mTBI in the chronic stage is a frequent referral fo the Veterans Health Administration. Conventional MRI and CT are typically normal months after civilian and military mTBI making it difficult to accurately diagnose and to determine rehabilitation strategies. Diffusion tensor imaging (DTI) can be used to characterize and quantify WM pathways in the living brain. Specific to brain injury, pathological processes causing loss or disorganization of fibers associated with breakdown of myelin and downstream nerve terminals, neuronal swelling or shrinkage, and increased or decreased extracellular space, could affect the quantitative scalar metrics like mean diffusivity (MD), fractional anisotropy (FA), radial diffusivity (RD), and/or axial diffusivity (AD). Recent studies have reported that FA was reduced in chronic civilian mTBI. Evidence from military cohorts also suggests important changes in DTI metrics across several brain regions. Machine learning (ML) algorithms are particularly sensitive to distributed changes caused by disease as observed in several structural and functional studies. This particular class of algorithms is specifically designed to identify patterns in temporal or spatial data to distinguish between groups. While several ML algorithms exists, one particular multivariate algorithm known as a Support Vector Machine (SVM) has been successfully applied to Alzheimer's Disease studies as well as a recent study in a group of TBI patients through the use of DTI data. In addition, the incorporation of principal component analysis (PCA) to SVM showed robust automated detection of WM degradation in Alzheimer's Disease over several sites and MR scanner platforms. This ability to evaluate this across platforms is particularly attractive to multi-center imaging studies that are performed in the VHA system. At present, the automated detection of biomarkers is scarce in the diagnosis and prognosis of mTBI in our Veteran population. This work will tailor an imaging and detection strategy that can possibly be used to not only identify Veterans with mTBI more objectively but also predict cognitive outcome to help facilitate appropriate rehabilitation strategies. Aim 1 will consist of a retrospective study of 70 subjects and controls to train the SVM algorithm to differentiate between mTBI pathology and uninjured military controls that were also deployed in the OIF/OEF/OND conflicts. DTI skeletons will be processed using Tract-Based Spatial Statistics (TBSS) software and will be used as inputs into the SVM algorithm. Using this data, parameters such as the cost function will be determined to optimize the algorithm. We will measure the accuracy, sensitivity and specificity of the algorithm by using a cross-validation approach. Finally for this first aim, we will use a sensitivity analysis techniqueto identify regions the algorithm weights more in determining if an mTBI has taken place. This will identify pathways that are vulnerable to injury. In Aim 2, we will use the SVM classifier on DTI scans to output possibility indices of mTBI. Regression analysis will be used to relate these indices to outcome measures. In conclusion, this work will provide a robust tool to not only better diagnose and characterize mTBI but also stratify more personalized rehabilitation strategies through the improved characterization of mTBI.