Helmet–Head Decoupling in Ice Hockey Impacts: An In-lab Exploratory Study Using Autoregressive Modeling of Acceleration Data Measured from a Helmet-Mounted Inertial Measurement Unit (IMU)

TL;DR

This study explores using AR modeling in helmet-mounted IMUs to better estimate head acceleration during ice hockey impacts, improving accuracy compared to raw or filtered data.

Contribution

A novel AR-based transfer function is proposed to mitigate helmet-head decoupling in impact measurements.

Findings

The AR-based transfer function reduced RMSE by up to 57% and MAPE by 27.54% compared to lowpass filtered signals.

Performance varied with helmet type, showing larger errors on a different helmet model.

The method consistently improved head acceleration estimation across all impact directions.

Abstract

This study aims to develop and validate in-lab a novel approach for estimating head linear acceleration in ice hockey impacts using IMU-instrumented helmets. The use of AutoRegressive (AR) modeling was investigated as a solution to mitigate the decoupling observed between the helmet and the head.

A series of impacts were conducted on a helmeted Hybrid III 50th percentile male Anthropometric Test Device (ATD). The impacts were performed using a custom-built pendulum impactor in four directions (front, front–oblique, side and back–oblique) and at two energies, 33 and 79 J, except for the back–oblique direction, which was tested only at 33 J. The processing pipeline included impact segmentation, main direction estimation and application of the AR-based transfer function modeling. The error with respect to the reference signals from the headform was quantified and the transformed signals were compared with the unprocessed (raw) and lowpass filtered signals. The generalization capabilities of the transfer function were also evaluated on a different helmet type.

The application of the transfer function resulted in a reduction of up to 9.04 g (57%) and 27.54% for the average Root Mean Squared Error (RMSE) and peak Mean Absolute Percentage Error (MAPE), respectively, with a consistent error decrease across all impact directions, compared to the lowpass filtered signal. However, when evaluated on a different helmet model, the transfer function showed larger errors.

The proposed methodology effectively improves the estimation of head linear acceleration across all impact directions. Nevertheless, performance varies with helmet type, indicating the need for helmet-specific adjustments (e.g., through model retraining).

The online version contains supplementary material available at 10.1007/s10439-025-03848-2.

Figures