K9 - On-field Head Acceleration Exposure Measurement Using Instrumented Mouthguards: Missing Data Imputation for Complete Exposure Analysis

Introduction: : Repetitive head acceleration events (HAEs) are increasingly recognized as contributors to both acute concussions and long-term neurological impairment in contact-sport athletes. While high-magnitude HAEs can result in immediate injury, there is growing concern that repetitive, lower-magnitude HAEs may also pose cumulative risks to brain health. However, the development of reliable exposure-risk models is hindered by inconsistent and incomplete HAE measurement. Instrumented mouthguards (iMGs) and video verification are widely used to quantify on-field HAEs, yet both methods are susceptible to missing data due to factors such as limited camera coverage, athlete noncompliance, and sensor malfunction. In our previous work, we observed substantial HAE data missingness despite adherence to current best practices. This study presents a novel approach to estimating complete HAE exposure by applying domain-specific imputation methods to address missing video and iMG data. We show that statistical mapping and machine learning-based multiple imputation (MI) can reconstruct full-season exposure profiles with greater accuracy than raw captured data alone. Our work demonstrates that missing data techniques are critical for generating complete and reliable HAE exposure estimates, thereby improving the quality of research into both acute and long-term risks of brain trauma in sport.

Materials and

Methods: : We monitored head acceleration exposure among university-level men’s ice hockey players throughout one varsity season using video footage and custom-fitted iMG sensors (Fig. 1A). Due to limited camera setups during away games, we used a statistical mapping approach to impute missing video-based HAE counts based on home game data (Fig. 1B i). We then addressed more substantial missingness in iMG data caused by sensor malfunction or noncompliance. Using a random forest-based MI approach, we predicted missing iMG HAE counts from observed video HAEs and additional variables, including athlete ID, game date, and game location (Fig. 1B ii). Imputed iMG HAEs were assigned timestamps from corresponding video-detected HAEs to allow for time-weighted exposure metric calculation. Kinematic values for each imputed HAE were then randomly sampled from a cross-verified pool of iMG events (n = 994). Final exposure estimates (including HAE counts, cumulative peak accelerations and brain strains) were obtained by pooling results across multiple imputed datasets (Fig. 1C).

Results, Conclusions, and Discussions:: Among 591 athlete-games, 45% had no iMG data due to noncompliance or malfunction. Imputation increased the median per-athlete-season video and iMG HAE counts by 10% and 69%, respectively (Table 1). Corresponding cumulative metrics, peak linear acceleration, peak angular acceleration, and peak corpus callosum strain, increased by 95%, 109%, and 69%, respectively. Leave-one-out cross-validation of the iMG HAE imputation model yielded a mean absolute error of 1.71 and a root mean squared error of 2.75 HAEs, demonstrating acceptable predictive performance.

This study introduces a novel two-stage imputation approach for estimating missing HAE data using both sensor and video sources. By quantifying iMG missingness at a full-season, per-athlete scale, we reveal the potential for severe underestimation in raw data. Our methods allow for per-game and per-season exposure estimation at the athlete level, enabling individualized risk profiling and more robust correlations between exposure and neurological outcomes.

The implications of missing HAE data are substantial, particularly for studies exploring the links between exposure and brain health. Without addressing missingness, exposure estimates may be biased, attenuating or distorting observed associations with clinical outcomes. Our approach helps mitigate these limitations, providing a more complete picture of athlete exposure. We recommend that future HAE studies integrate similar imputation strategies to enhance data integrity, support individualized assessments, and improve the strength of longitudinal analyses in sports concussion research. Our future work will continue to optimize our imputation models for greater prediction capabilities.

Acknowledgements and/or References (Optional)::