Sensitivity of Brain Injury Models to Head Pose, Sample Rate and Interpolation

Purpose: This study investigates how the lower sampling rates and interpolation methods typical of photogrammetric processes for head impact kinematics measurement can affect the accuracy of brain strain predictions, a critical metric in understanding traumatic brain injury (TBI) and mild traumatic brain injury (mTBI). Methods: A dataset of head impact kinematics was generated using emulated instrumented mouthguard (iMG) data sampled at 1000 Hz. These data were converted to head pose measurements and downsampled to rates reflective of video-based systems (500, 100 and 50 Hz). Interpolation schemes (linear, cubic spline and PCHIP) were applied to upsample the downsampled data back to 1000 Hz before recovering angular velocity. Brain strain, measured as the 95th percentile of Maximum Principal Strain (MPS95), was predicted using both a traditional finite element head model (the EdiFEHM) and a convolutional neural network (CNN). Results: Interpolation marginally improved the accuracy of strain predictions at lower sample rates, with PCHIP performing best among the methods tested. However, even with interpolation, over 25% of predictions at 50 Hz deviated by more than 10% from the 1000 Hz reference. The CNN displayed heightened sensitivity to linear interpolation compared to the EdiFEHM, due to its inability to account for the spikes present in the piecewise continuous angular velocity profiles. Conclusion: Higher sample rates (≥ 500 Hz) yield accurate brain strain predictions, but lower rates (e.g., 50 Hz) introduce significant inaccuracies, despite marginal improvements with simple polynomial interpolation.