Predicting the Internal Knee Abduction Impulse During Walking Using Deep Learning.

Area of Science:

• Biomechanics
• Machine Learning
• Sports Science

Background:

• Knee joint moments estimate knee joint loads but inverse dynamics methods are time-consuming.
• Developing efficient methods to calculate knee joint loads is crucial for research and clinical applications.

Purpose of the Study:

• To benchmark five deep learning models for predicting internal knee abduction impulse during walking using segment kinematics.
• To compare deep learning approaches against traditional inverse dynamics for quantifying knee abduction impulse.

Main Methods:

• Utilized a publicly available 3D kinematic and kinetic dataset of walking (n=33).
• Derived 126 time-series predictors from lower body segment kinematics (displacement, velocity, acceleration).
• Trained and tested five deep learning models, including a baseline 2D convolutional network and InceptionTime with transfer learning.

Main Results:

• Transfer learning with InceptionTime achieved the lowest mean absolute percentage error (MAPE) of 8.28%.
• A baseline 2D convolutional network model had a MAPE of 10.80%.
• Time-series based deep learning models outperformed an image-based deep learning approach (MAPE 16.17%).

Conclusions:

• Deep learning, particularly time-series models like InceptionTime with transfer learning, can accurately predict knee abduction impulse during walking.
• These findings suggest deep learning offers a more efficient alternative to inverse dynamics for estimating knee joint loads.
• Optimal network architectures and the benefits of transfer learning are key for developing wearable technologies for joint moment prediction.