Using Machine Learning techniques for identification of Chronic Traumatic Encephalopathy related Spectroscopic Biomarkers

Abstract

Contact sports athletes, military personnel, and civilians that suffer from multiple head traumas have the potential to develop Chronic Traumatic Encephalopathy (CTE), a progressive, degenerative brain disease diagnosed only postmortem by characteristic tau deposition in the brain. There is, therefore, a need for in-vivo diagnosis for CTE to diagnose and manage this disease, while the individual is still alive. However, there is no definitive in-vivo diagnosis because of heterogeneous clinical symptoms that often overlap with other neurodegenerative diseases. Magnetic Resonance Spectroscopy (MRS) can be a suitable candidate for CTE diagnosis as multiple head trauma changes the neurochemicals in the brain that can be detected using MRS. These changes can be subtle, and group differences are not sufficient for clinical diagnosis. This paper proposes a machine learning based approach to capture the neuro-spectroscopic signatures corresponding to CTE-related impairments in NFL players. The classification model uses concentration estimates of metabolites to classify between ‘Impaired and ‘Non-impaired players. The model using the metabolite concentrations of creatine, choline, N-acetyl-aspartate, glutamate, and macromolecules achieved Area Under the Curve (AUC) of 0.72 and prediction accuracy of 75%. While these metabolites have been shown to be altered in previous concussion studies, other metabolites may improve the diagnostic accuracy. In order to include more metabolites, two-dimensional correlated spectroscopy (L-COSY), which resolves overlapping metabolites, was also acquired. The L-COSY model which included 15 metabolites, increased prediction accuracy to 87 % with AUC of 0.83. With the aid of machine learning, these metabolites may serve as potential biomarkers that correspond to the CTE-related impairments that will allow for CTE diagnostics in athletes prior to their death.