Machine learning techniques for estimating the individual three-dimensional ground reaction forces during rugby...

Area of Science:

• Sports Science
• Biomechanics
• Machine Learning

Background:

• Rugby scrummaging is crucial for match outcomes, driven by forward horizontal force.
• Current methods for measuring scrummaging forces are limited, necessitating portable solutions.
• Instrumented insoles can measure ground reaction forces (GRF), but face challenges with non-vertical forces.

Purpose of the Study:

• To compare four machine learning algorithms for estimating 3D GRF during rugby scrummaging using instrumented insoles.
• To evaluate the effectiveness of dataset expansion and model personalization for improved accuracy.
• To determine the best model for accurately assessing individual player force contributions in live scrums.

Main Methods:

• Four machine learning algorithms were evaluated: Random Forest, Multi-Layer Perceptron (MLP), Long-Short-Term Memory (LSTM), and a hybrid LSTM-MLP.
• The study utilized instrumented insoles to capture GRF data during rugby scrummaging.
• Training configurations included dataset expansion and model personalization.

Main Results:

• A personalized MLP model trained on a reduced dataset achieved the best performance.
• Normalized root mean square errors were 1.8% (Medio-Lateral), 5.6% (Antero-Posterior), and 8.3% (Vertical).
• A non-personalized MLP trained on an expanded dataset also showed high accuracy, suitable for new individuals.

Conclusions:

• Machine learning, particularly personalized MLP, can accurately estimate 3D GRF from instrumented insoles in rugby scrummaging.
• This technology enables field-based assessment of individual player force contributions.
• The findings support the use of instrumented insoles and ML for biomechanical analysis in team sports.