Predicting 3D ground reaction forces across various movement tasks: a convolutional neural network study comparing...

Area of Science:

• Biomechanics
• Machine Learning
• Wearable Technology

Background:

• Ground reaction forces (GRFs) are essential for analyzing movement and musculoskeletal load.
• Inertial measurement units (IMUs) enable gait analysis in real-world settings but cannot directly measure GRFs.
• Machine learning offers a promising approach to estimate 3D-GRFs from IMU data, yet research has primarily focused on vertical GRF and limited movement types.

Purpose of the Study:

• To systematically evaluate the prediction accuracy of convolutional neural networks (CNNs) for 3D-GRFs.
• To assess prediction performance using IMUs from various sensor configurations (single and multiple).
• To analyze accuracy across diverse movement tasks, including walking, stair negotiation, and cutting maneuvers.

Main Methods:

• Twenty healthy participants performed six distinct movement tasks at self-selected speeds.
• CNNs were trained to predict 3D-GRFs using IMU time-series data from configurations including lower body (7 IMUs), single leg (4 IMUs), femur-tibia (2 IMUs), tibia (1 IMU), and pelvis (1 IMU).
• Prediction accuracy was evaluated using leave-one-subject-out cross-validation, employing Pearson correlation (r) and relative root mean squared error (relRMSE).

Main Results:

• CNNs demonstrated high accuracy in predicting vertical GRF (vGRF) (r=0.98, relRMSE ≤ 7.44%) across all tasks.
• Anterior-posterior GRF (apGRF) prediction was also accurate (r ≥ 0.92, relRMSE ≤ 14.24%), while medial-lateral GRF (mlGRF) prediction was less accurate (r ≥ 0.74, relRMSE ≤ 29.46%).
• vGRF prediction accuracy was consistent between multi-IMU and single-IMU configurations, supporting the use of simpler setups for vGRF estimation. Multi-IMU configurations showed improved accuracy for mlGRF and apGRF during cutting maneuvers.

Conclusions:

• CNNs effectively predict 3D-GRFs from IMU data across various movements.
• Single IMU configurations are sufficient for accurate vGRF prediction, offering practical advantages.
• Advanced sensor configurations (e.g., lower body) enhance prediction accuracy for apGRF and mlGRF during complex dynamic tasks like cutting maneuvers, indicating a trade-off between sensor simplicity and predictive performance.