Lower Limb Activity Recognition Based on sEMG Using Stacked Weighted Random Forest.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Surface electromyography-based pattern recognition systems (sEMG-PRS) face challenges with generalizability in real-world use.
• Existing systems struggle with long-term usability and adapting to individual users.

Purpose of the Study:

• To introduce a novel stacked weighted random forest (SWRF) algorithm.
• To enhance the generalizability, long-term usability, and user adaptability of sEMG-PRS.

Main Methods:

• Developed a weighted random forest (WRF) to address performance imbalance in standard random forests (RF).
• Employed stacking, using RF as base learners and WRF as the meta-learner, to create the SWRF algorithm.
• Evaluated SWRF against RF, WRF, LSTM, and SVM using offline datasets and online experiments.

Main Results:

• SWRF achieved an average offline classification accuracy of 89.06%, surpassing RF, WRF, LSTM, and SVM.
• Online experiments demonstrated SWRF's superior performance in long-term usability and user adaptability compared to other algorithms.

Conclusions:

• The proposed SWRF algorithm significantly improves sEMG-PRS performance.
• SWRF shows strong potential for practical applications requiring robust and adaptable human-computer interfaces.