A neural network method to predict task- and step-specific ground reaction force magnitudes from trunk accelerations...

Area of Science:

• Biomechanics
• Sports Science
• Machine Learning

Background:

• Accurate prediction of ground reaction forces (GRF) is crucial for optimizing training loads and preventing injuries in athletes.
• Current methods often rely on multiple sensors, limiting practical application, or estimate only discrete parameters, failing to capture overall biomechanical load.

Purpose of the Study:

• To develop and validate a novel neural network-based method for predicting GRF time series.
• To utilize data from a single, trunk-mounted accelerometer for GRF prediction, enhancing ecological validity.

Main Methods:

• A principal component analysis (PCA) combined with a multilayer perceptron (MLP) was employed to predict GRF time series.
• The model was trained and tested using data from trunk accelerometers during running-based activities.

Main Results:

• The neural network achieved high accuracy in predicting GRF time series, with average time-series r² coefficients around 0.9.
• Impact peak prediction yielded an r² of 0.74 across various activities, outperforming traditional correlation methods.
• The method demonstrated effective GRF estimation from a single accelerometer, negating the need for additional sensors.

Conclusions:

• The proposed neural network method accurately predicts GRF time series from trunk accelerometry data.
• This approach offers a viable, sensor-efficient solution for biomechanical load assessment in sport-specific environments.
• Machine learning holds significant potential for leveraging common wearable technology in sports performance and injury prevention.