Self-supervised learning enhances accuracy and data efficiency in lower-limb joint moment estimation from gait...

Area of Science:

• Biomechanics
• Deep Learning
• Machine Learning

Background:

• Estimating human joint moments is crucial for assessing joint loads but traditionally requires extensive labeled data.
• Collecting synchronized joint angle and moment data is challenging in real-world biomechanical applications.
• Deep learning (DL) offers potential but faces data limitations for accurate joint moment estimation.

Purpose of the Study:

• To improve the accuracy and data efficiency of knee joint moment estimation.
• To leverage self-supervised learning (SSL) for extracting human motion representations from unlabeled data.
• To reduce the reliance on large labeled datasets for training joint moment estimation models.

Main Methods:

• A Transformer auto-encoder architecture was employed for SSL-based joint moment estimation.
• The model was pre-trained on large unlabeled joint angle datasets using masked reconstruction.
• Fine-tuning was performed on a small subset of labeled joint moment data.

Main Results:

• The SSL model significantly outperformed the baseline, especially with limited labeled data (e.g., 5% labeled data).
• Mean Squared Errors (MSE) and Mean Absolute Errors (MAE) were substantially reduced compared to the baseline.
• The SSL model achieved superior performance even when using only 20% of the labeled data compared to the baseline using 100%.

Conclusions:

• Self-supervised learning significantly enhances the accuracy and data efficiency of joint moment estimation.
• The proposed method offers a more efficient solution for biomechanical evaluations, reducing data collection burdens.
• This approach has the potential to expand the clinical applicability of joint moment estimation techniques.