Feature Pruning in Deep Neural Networks for Gait Speed Prediction: Evaluating Performance Without Domain-Specific
Summary
Feature pruning in deep neural networks significantly improves gait speed prediction accuracy using electromyography data. This method reduces computational costs and enhances model performance, making AI-driven gait analysis more accessible for clinical applications.
Area of Science:
- Biomedical Engineering
- Machine Learning in Healthcare
- Neuromuscular Function Analysis
Background:
- Gait speed is a key indicator of mobility and neuromuscular health.
- Deep learning models can predict gait speed from electromyography (EMG) data.
- Extensive feature sets in deep learning may increase complexity without improving accuracy.
Purpose of the Study:
- To investigate the impact of feature pruning on deep neural network (DNN) models for gait speed prediction.
- To evaluate performance using only frequency-domain muscle activation features.
- To assess if feature reduction enhances accuracy and efficiency without domain-specific feature selection.
Main Methods:
- A deep neural network (DNN) was trained on frequency-domain EMG features.
- A group-wise sequential feature selection method reduced 540 features to 60.
- Model performance was evaluated based on root mean square error (RMSE) and computational cost.
Main Results:
- Feature pruning reduced training time by 79% and computational cost by 81.4%.
- The pruned model improved RMSE from 0.793 km/h to 0.610 km/h.
- The pruned DNN outperformed the full-featured model across all gait speeds.
Conclusions:
- Systematic feature reduction enhances DNN model performance and generalizability for gait speed prediction.
- Feature pruning offers a viable approach for AI-driven gait analysis in clinical settings.
- This method facilitates the deployment of predictive models on wearable devices for continuous monitoring.
