Feature Pruning in Deep Neural Networks for Gait Speed Prediction: Evaluating Performance Without Domain-Specific
Abstract:
Gait speed is a critical biomarker for assessing mobility, rehabilitation, and neuromuscular function. While deep learning (DL) models have shown promise in predicting gait speed using electromyography (EMG) data, including extensive feature sets can increase computational complexity without necessarily improving accuracy. With 17 discrete gait speeds and 35 participants' gait data, this study explores the impact of feature pruning in a deep neural network (DNN) trained only on frequency-domain muscle activation features, excluding domain-specific knowledge in the feature selection process. A group-wise sequential feature selection method was applied to reduce the original 540 features to 60. The test results demonstrate that feature pruning significantly reduced training time (by 79%) and computational cost (81.4% fewer parameters) while improving root mean square error (RMSE) from 0.793 km/h to 0.610 km/h. Despite excluding domain-specific guidance, the pruned model performed better than the full-featured DNN at all gait speeds. These findings suggest that systematic feature reduction can enhance model performance and generalizability, even without expert-driven feature selection. Future studies should examine the approach across diverse populations and gait conditions to validate its applicability.Clinical relevance- Accurate gait speed prediction is crucial for diagnosing and monitoring neuromuscular disorders, assessing fall risk, and evaluating rehabilitation progress. This study demonstrates that deep learning can be used in gait speed predictions even without expert-driven feature selection, making AI-driven gait analysis more accessible in real-world clinical settings. Feature pruning will facilitate the deployment of predictive models on portable and wearable devices for continuous gait monitoring and early detection of mobility impairments.
