Multidimensional Ground Reaction Forces and Moments From Wearable Sensor Accelerations via Deep Learning.

Area of Science:

• Sports Science
• Biomechanics
• Machine Learning

Background:

• Monitoring athlete internal workload is crucial for preventing non-contact knee injuries.
• Current methods for measuring musculoskeletal joint loads are laboratory-bound, costly, and lack real-world applicability.

Purpose of the Study:

• To develop a novel method for obtaining ground kinetics in field settings.
• To estimate accurate, reliable, and valid musculoskeletal joint loads in near real-time during athletic activities.

Main Methods:

• Utilized supervised learning techniques, specifically convolutional neural network (CNN) deep learning models.
• Trained models using laboratory-derived ground reaction forces and moments (GRF/M) data with simulated accelerometer data from extensive motion trials.
• Validated predictions using sensor accelerations recorded during independent inter-laboratory data capture sessions.

Main Results:

• Achieved high correlations between predicted and ground truth GRF/M for vertical (0.97) and anterior (0.96) components during running.
• Demonstrated moderate correlations for lateral GRF (0.87) during sidestepping and for GRM (0.65).
• The best-case correlations indicate the approach's plausibility, though results showed variability.

Conclusions:

• The deep learning approach shows potential for estimating on-field ground kinetics using wearable sensors.
• Lessons learned will inform future improvements for accurate near real-time on-field GRF/M estimation.
• This technology could aid in monitoring joint loading to minimize non-contact injuries in sports.