Comparing shallow, deep, and transfer learning in predicting joint moments in running

Area of Science:

• Biomechanics
• Machine Learning
• Sports Science

Background:

• Joint moments are crucial for understanding muscle behavior and joint loading in biomechanics.
• Quantifying joint moments clinically or in the field is challenging due to difficulties in measuring ground reaction forces outside laboratory settings.

Purpose of the Study:

• To compare the accuracy of three machine learning (ML) techniques—functional regression (MLfregress), a deep neural network (MLDNN), and transfer learning (MLTL)—in predicting joint moments during running.
• To evaluate the potential of ML models using kinematic data to overcome current biomechanics data acquisition constraints.

Main Methods:

• Utilized an open-source dataset and data from two running studies with and without external loads.
• Derived 3D joint moments (hip, knee, ankle) via inverse dynamics.
• Employed 3D joint angle, velocity, and acceleration as predictors for MLfregress, an ML DNN built from scratch, and MLTL.

Main Results:

• Prediction performance was highest with ML_DNN and lowest with ML_fregress.
• The best absolute predictive performance was observed for sagittal plane moments, with RMSE values ranging from 0.16 Nm/kg (ankle, ML_DNN) to 0.49 Nm/kg (knee, ML_fregress).
• ML_DNN demonstrated a 20% improvement in relative prediction performance (relRMSE) for ankle adduction-abduction moments compared to ML_fregress.

Conclusions:

• Deep neural networks (DNNs), with or without transfer learning, significantly outperformed functional regression in predicting joint moments using kinematic inputs.
• Integrating ML with kinematic data offers a promising solution for obtaining high-fidelity biomechanics data outside traditional laboratory environments.