Rugby scrummaging represents a critical phase of play, with its outcomes closely associated with overall match performance. Scrum success is primarily determined by the forward-directed horizontal force generated by the entire pack. Previous studies have focused either on the total force produced by the pack or on individual efforts against a scrum machine. To evaluate the contribution of each player to horizontal force production within a live scrum, portable measurement systems are required. Instrumented insoles offer a field-based solution for measuring ground reaction forces (GRF). However, they are generally limited to measuring only the perpendicular component of force to the insole surface, for situations involving mainly vertical GRF, with flat-foot contact. The objective of this study was to compare the performance of four Machine Learning algorithms (Random Forest, Multi-Layer Perceptron (MLP), Long-Short-Term Memory, and a combination of a Long-Short Term Memory and an MLP) for estimating the three-dimensional GRF during rugby scrummaging using instrumented insoles. Various training configurations were tested, including dataset expansion through the merging of two experimental datasets and the application of model personalization. The best results were obtained using a personalized MLP trained on the reduced dataset, yielding root mean square error values normalized by body weight of 1.8 ± 0.3 % (Medio-Lateral), 5.6 ± 1.1 % (Antero-Posterior), and 8.3 ± 2.2 % (Vertical). The non-personalized MLP model trained on the extended dataset also demonstrated strong performance, indicating its suitability for application to new individuals with minimal reduction in accuracy.
© Copyright 2025 Journal of Biomechanics. Elsevier. Alle Rechte vorbehalten.