74. blast-related mild traumatic brain injury in chronic phase: a diffusion tensor imaging study with machine learning

Department: Computer Science & Engineering
Faculty Advisor(s): Chung K. Cheng

Primary Student
Name: Poya Hsu
Email: p8hsu@ucsd.edu
Phone: 858-534-8174
Grad Year: 2022

Individuals exposed to blast-induced mild traumatic brain injury (bmTBI) often experience long-term neurological and cognitive disorders. Until now, no optimal medical treatments have been developed specifically for bmTBI. However, a well-designed neuroimaging exper- iment with appropriate analyses would aid the understanding of bmTBIs? mechanisms and treatment targets. In this study, we recruited deployed service members with combat ex- posure, 19 healthy controls (HC) with no bmTBI and 20 active-duty service members or Veterans with chronic bmTBI. Diffusion tensor imaging data and neuropsychological (NP) performance data were collected, and voxel-wise statistical analyses were carried out using Tract-Based Spatial Statistics (TBSS). Subsequently, the data attained from TBSS were utilized as features in tract-based and data-driven models in support vector machines for classification. TBSS showed relatively increased Fractional Anisotropy (FA) and relatively reduced Radial Diffusivity (RD) in bmTBI compared to HC. Furthermore, tract-based re- sults based on FA and NP correctly classified bmTBI versus controls in 74% and 77% of cases respectively. To boost the classification performance, we came up with a novel feature by making use of the metrics of DTI results: axial diffusivity (AD) and radial diffusivity (RD) in cross-validation. The new results reached 94% accuracy solely using FA and AD/RD experiments. A model of the mixture of FA, AD/RD , and NP scores reached 96% accuracy in classification.

Industry Application Area(s)
Software, Analytics | Neuroimaging

« Back to Posters or Search Results