Performance of deep-learning models incorporating knee alignment information for predicting ground reaction force...

Area of Science:

• Biomechanics
• Machine Learning
• Wearable Technology

Background:

• Deep learning models and wearable sensors are advancing biomechanical variable prediction.
• Integration of knee alignment and injury-side data is often overlooked in predictive models.
• Previous research has explored simple neural networks and complex pretrained models for predicting 3D ground reaction force (GRF).

Purpose of the Study:

• To compare the performance of various deep learning model architectures, including complex pretrained models, in predicting 3D GRF during level walking.
• To evaluate the impact of incorporating subject-specific features, such as knee alignment, on prediction accuracy.
• To identify the most efficient and accurate model for real-world biomechanical analysis.

Main Methods:

• Developed ten deep-learning models using motion capture and wearable accelerometer data.
• Compared models with and without subject-specific features, including knee alignment and injury side.
• Evaluated established pretrained models (ResNet50, Inception) and custom hybrid models (2D-CNN-LSTM).

Main Results:

• Incorporating subject-specific features improved prediction accuracy for most models, except LSTM.
• A 2D-CNN-LSTM hybrid model demonstrated the highest prediction accuracy.
• Pretrained models (ResNet50, Inception) benefited from ImageNet weights and subject-specific features but did not surpass custom hybrid models for time-series GRF prediction.
• Larger pretrained models showed increased computational demands without superior time-series performance.

Conclusions:

• Subject-specific features, particularly alignment information, significantly enhance the accuracy of walking GRF predictions.
• Custom hybrid deep learning models outperform complex pretrained models for time-series 3D GRF prediction.
• Custom models integrating alignment features offer a more efficient and effective solution for accurate, resource-constrained biomechanical predictions.