Related Experiment Video
Updated: Jan 15, 2026
An Inertial Measurement Unit Based Method to Estimate Hip and Knee Joint Kinematics in Team Sport Athletes on the Field
Published on: May 26, 2020
Self-supervised learning enhances accuracy and data efficiency in lower-limb joint moment estimation from gait
Yifan Li1,2, Jiayu He1, Bernard Liew3
1Department of Engineering, King's College London, London, United Kingdom.
Objective:
Deep learning (DL) has introduced new possibilities for estimating human joint moments - a surrogate measure of joint loads. However, traditional methods typically require extensive synchronised joint angle and moment data for model training, which is challenging to collect in real-world applications. This study aims to improve the accuracy and data efficiency of knee joint moment estimation via leveraging self-supervised learning techniques to automatically extract human motion representations from large-scale unlabeled joint angle datasets.
Method:
We proposed a joint moment estimation method based on self-supervised learning (SSL), using a Transformer auto-encoder architecture. The model was pre-trained on large-scale unlabeled joint angle data with masked reconstruction to effectively capture spatiotemporal features of human motion. Subsequently, we fine-tuned the model using a small amount of labeled joint moment data, enabling accurate mapping from joint angles to joint moments. We evaluated this method on a dataset of 55 normally developing children and compared the performance of the pre-trained SSL model fine-tuned with different amounts of labeled data to a baseline model.
Results:
The Fine-tuned model significantly outperformed the baseline model, especially in scenarios with scarce labeled data. MSEs were reduced from 24.00% to 45.16% (with an average reduction of 36.29%), and MAE from 18.18% to 37.80% (with an average reduction of 26.48%). The proposed SSL model exceeded the performance of the baseline model trained with 100% data, using only 20% of the data in the labeled dataset during fine-tuning. When both models were fine-tuned using only 5% of the labeled data, the proposed SSL achieved four-fold better performance than the baseline model.
Conclusion:
This study demonstrates that self-supervised learning significantly improves the accuracy and data efficiency of joint moment estimation, providing a more efficient solution for biomechanical evaluation. The proposed model can reduce the burden of collecting data and expand clinical applications.
