Accurate estimation of brain tissue strain in sports is often limited by the high cost of sensors and the computational demands of finite element simulations. This study presents a deep-learning framework to predict maximum principal brain strain directly from video-derived biomechanical features and player age in youth ice hockey. We analysed 477 on-ice head impacts involving players aged 5-18 years, with reference strain values obtained from a previously validated finite element brain model. For each impact, four features, player velocity, surface compliance, impact location, and elevation, plus player age group were extracted. A fully connected neural network was trained using a five-run repeated hold-out protocol with 75%, 15%, and 10% of impacts used for training, validation, and testing, respectively, in each run. On the held-out test sets, the model achieved a mean squared error of 0.00142 ± 0.00051 and a mean coefficient of determination of 0.862 ± 0.032, where values represent mean ± standard deviation across runs. Permutation importance analysis identified impact velocity as the dominant predictor, followed by surface compliance, whereas location, elevation, and age contributed less within this dataset. The model converged quickly and required only about 8 ms per prediction on a standard desktop computer, indicating that inference is computationally lightweight. Although the present study was conducted offline, the same approach could be integrated into an automated video-analysis pipeline for event detection, head-impact detection, feature extraction, and velocity estimation, enabling near real-time assessment. This video-based, non-invasive method provides a scalable solution for monitoring brain strain in youth hockey and could potentially be extended to other sports.
© Copyright 2026 Sports Engineering. The Faculty of Health & Wellbeing, Sheffield Hallam University. Alle Rechte vorbehalten.