Using a pre-trained machine learning model to estimate the 3d-ground reaction forces during rugby scrummaging with instrumented insoles
Résumé
INTRODUCTION:Rugby scrummaging represents a crucial phase of the game, characterized by high-intensity physical efforts and a significant impact on match outcomes [1]. The horizontal force generated by the entire pack is a key determinant of scrum success. However, existing measurement systems are unable to provide 3D, individual, and on-field assessments of ground reaction forces (GRF). A previous study developed a Machine Learning (ML) model to predict the 3D-GRF with instrumented insoles during scrummaging but this was conducted on recreationally active subjects without specific scrummaging experience [2]. Thus, this study aimed to investigate to what extent this model can be used to predict the 3D-GRF for elite rugby players.METHODS:Twelve elite rugby players (12 males; age: 20+/-1 ans; height: 191+/-7 cm; weight: 116+/-13 kg) performed three pushing trials of 15 seconds, against a fixed scrum machine. They wore commercial instrumented insoles (Loadsol Pro®, Novel, Germany, 200Hz) inside their shoes with each foot on a force plate (Sensix, 1000Hz) covered with artificial turf. The force plate data served as the reference for 3D-GRF measurements. For each subject, one trial was used to infer data from the pre-trained ML model, while the remaining two trials were utilized to personalize the model for each subject. The model’s performance was assessed by computing the Root Mean Square Error (RMSE) between the prediction and the reference, the correlation coefficient (r), and the percentage of RMSE compared to the mean resultant force.RESULTS:The initial model inference yielded mean RMSE values of 42±9N on the Medio-Lateral (ML) axis, 168±87N on the Antero-Posterior (AP) axis, and 180±49N on the Vertical (V) axis, with correlation coefficients r of 0.708±0.110 (ML), 0.825±0.102 (AP), and 0.571±0.123 (V). RMSE percentages relative to the mean resultant force were 4.2±0.7% (ML), 15.8±5.0% (AP), and 18.0±4.8% (V). After personalization, RMSE were 32±10N (ML), 99±41N (AP), and 135±38N (V), and correlation coefficients r were0.838±0.062 (ML), 0.911±0.044 (AP), and 0.680±0.168 (V). The mean percentages of RMSE compared to the average resultant force were 3.1±0.6% (ML), 9.7±2.8% (AP), and 14.0±5.5% (V).CONCLUSION:The findings of this study indicate that an ML model pre-trained on data from recreationally active individuals without specific rugby scrummaging experience is not optimal for accurately estimating 3D-GRF in elite rugby players. Model personalization for each participant improved performance, suggesting that personalization is a promising approach for enhancing ML model performance when the model is trained on a non-specific dataset. However, further improvement in model performance could be achieved by pre-training the model on a dataset more closely aligned with the data used for personalization, to ensure reliable predictions.References[1] Scott et al., J. Sci. Med. Sport, 2023[2] Pomarat et al., IEEE Xplore, 2025