Bayesian Spatial Models for Predicting the Location of Head Impacts

• Activity Number: 368
• Type: Contributed
• Date/Time: Tuesday, August 2, 2016 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Michael Lawson* and Daniel Hernandez-Stumpfhauser and Amy Herring and Gunter Siegmund and Jason Mihalik and Steve Marshall and Kevin Guskiewicz
• Companies: The University of North Carolina at Chapel Hill and The University of North Carolina at Chapel Hill and The University of North Carolina at Chapel Hill and MEA Forensic and The University of North Carolina at Chapel Hill and The University of North Carolina at Chapel Hill and The University of North Carolina at Chapel Hill
• Keywords: Bayesian modeling ; spatial statistics ; traumatic brain injury ; projected normal distribution

Abstract

Effectively measuring head impacts is a crucial step in detecting and preventing traumatic brain injury, which affects millions of people across the world every year. Although accelerometer devices report the location and magnitude of head impacts, these devices give imperfect output. Their lack of accuracy and precision limits both the utility of the devices as on-field clinical tools and the validity of the data they produce for analysis. We develop a Bayesian spatial model to predict the true direction of the peak linear acceleration vector of a head impact based on the device-reported direction and magnitude of peak linear and angular acceleration-information available for applications of our model to prediction in real time. We fit the model via Gibbs sampling, drawing the angular outcome from the projected normal distribution and using slice sampling to draw the latent vector lengths. We assess the model's performance through simulation and apply the model to experimental data from a recent football helmet study. We furthermore demonstrate that the spatial dependencies built into our model improve its flexibility through a hold-one-out analysis of twelve key impact locations.