Multi-Model Gait Fatigue Detection and Feature Sensitivity Analysis Based on the DUO-GAIT Dataset

This study focuses on gait-based fatigue detection using the DUO-GAIT dataset. Through a rigorously designed experimental protocol, we conducted a comprehensive performance comparison of five computational models: MLP, BiLSTM, XGBoost, SVM, and Random Forest, complemented by an in-depth sensitivity analysis of feature discriminative power for fatigue classification. The experimental results demonstrate that the BiLSTM model achieves superior performance in fatigue state recognition, with stride times and stride lengths emerging as the most discriminative features for fatigue detection. In addition, ensemble learning models, such as XGBoost exhibit significant advantages in inference efficiency and model size, making them well-suited for practical deployment. SVM offers a balance between inference speed and model compactness, also showing potential for real-world applications. While deep learning models-especially BiLSTM-excel in recognition accuracy, they are less competitive in terms of inference speed and model size. These findings provide a critical theoretical foundation and practical guidelines for developing wearable device-based gait monitoring systems capable of continuous fatigue assessment in free-living environments. However, this study does not provide an in-depth quantitative analysis of other practical constraints for real-world wearable deployment, such as energy consumption, battery life, device computational resources, or whether the inference latency meets real-time application requirements. Future work will address these aspects to further facilitate efficient deployment of the proposed models on wearable devices. (Abstract)