AI-Based Denoising of Head Impact Kinematics Measurements With Convolutional Neural Network for Traumatic Brain Injury Prediction
Abstract
Wearable devices used for measuring head impact kinematics are inherently noisy due to imperfect coupling with the human body. This study developed one-dimensional convolutional neural network (1D-CNN) models to denoise tri-axial linear acceleration and angular velocity signals recorded from instrumented mouthguards. Using 163 dummy head impacts for training, validation was performed at three levels: kinematics, brain injury criteria, and tissue-level strain and strain rate. Denoising reduced pointwise RMSE by 36% and peak absolute error by 56%, while six brain injury criteria errors decreased by 82% on average. Maximum principal strain and strain-rate errors were reduced by 35% and 69%, respectively. Blind testing on 118 college football impacts and 413 PMHS impacts demonstrated similar improvements. The 1D-CNN approach provides an effective method for reducing measurement noise in mouthguard-derived head kinematics and supports future real-world TBI monitoring.
