Self-Supervised Learning Improves Accuracy and Data Efficiency for IMU-Based Ground Reaction Force Estimation.

Area of Science:

• Biomechanics
• Machine Learning
• Wearable Technology

Background:

• Deep learning models for inertial measurement unit (IMU)-driven kinetic assessment typically require extensive ground reaction force (GRF) data for supervised training.
• This reliance on labeled GRF data presents a significant bottleneck for practical applications.

Purpose of the Study:

• To investigate the efficacy of self-supervised learning (SSL) techniques for pre-training deep learning models using large IMU datasets.
• To improve the accuracy and data efficiency of IMU-based GRF estimation.

Main Methods:

• SSL was performed by masking portions of IMU data and training a transformer model to reconstruct the masked segments.
• The study compared various masking ratios across datasets of real, synthetic, and combined IMU data.
• Models were then fine-tuned with labeled data for GRF estimation in overground walking, treadmill walking, and drop landing tasks.

Main Results:

• SSL pre-training significantly enhanced the accuracy of 3-axis GRF estimation during walking compared to conventional supervised learning.
• Fine-tuning SSL models with only 1-10% of labeled walking data achieved accuracy comparable to baseline models trained with 100% of the data.
• Optimal masking ratios for SSL were identified as 6.25-12.5%.

Conclusions:

• SSL effectively utilizes large IMU datasets (real and synthetic) to boost the accuracy and data efficiency of deep learning-based GRF estimation.
• This approach substantially reduces the requirement for labeled GRF data, making IMU-driven kinetic assessment more accessible.