Traumatic brain injuries (TBIs) affect millions of people in the U.S., whether from car accidents, on the battlefield, or from sports injuries. Yet treatment and, in many cases, even a detailed diagnosis remains rather rudimentary. Outcomes can range from a complete recovery after seemingly severe injuries to debilitating depression or other personality disorders from apparently much milder incidents.

LLNL is collaborating with the TRACK-TBI Consortium—by far the largest effort of its kind—aimed at collecting comprehensive imaging, clinical, and outcome data for thousands of TBI patients across multiple U.S. emergency rooms. We hope this unprecedented collection of data will lead to fundamental new insights into how to diagnose and ultimately treat TBI.

While thousands of patients represent an order of magnitude larger data collection than in any previous study, the dataset is minuscule in the world of deep learning. Furthermore, the complex graph structures resulting from imaging data cannot be treated as images, nor do they resemble social networks or natural languages. To tackle these challenges, CASC is working on new ML approaches to graph representation learning, multimodal data analysis, small data techniques, and interpretable learning.

model of the structural connectivity of the human brain

TRACK-TBI and the DOE Mission

Understanding the impact of a brain injury and its possible treatments will push the boundaries and capabilities of scientific ML at national labs.

Bill Goldstein (right) shakes hands with San Francisco 49ers chair John York at a meeting on traumatic brain injury research at LLNL

NFL and TBI Research

Officials from the National Football League visited LLNL recently to hear about advancing scientific understanding of TBI with HPC and AI.

The T1 scans are parcellated to 84 different regions in the brain, on which, tractography is performed to compute the weighted matrix representing the structural connectome.

Modeling Human Brain Connectomes

The Human Connectome Project meets a structured network architecture to predict meta-information.