Adaptive Lower Limb Pattern Recognition for Multi-Day Control.

Area of Science:

• Biomedical Engineering
• Rehabilitation Engineering
• Human-Computer Interaction

Background:

• Pattern recognition in electromyography (EMG)-based control systems faces performance degradation due to concept drift.
• Concept drift, caused by factors like fatigue and sensor variability, leads to increased error rates in myoelectric control.
• Adaptation strategies are crucial for retraining pattern recognition systems to maintain performance over time.

Purpose of the Study:

• To investigate the development of error rates in EMG-based lower limb pattern recognition over one week.
• To compare the effectiveness of three distinct adaptation strategies in mitigating concept drift.
• To evaluate entropy-based adaptation as a potential solution for improving long-term myoelectric control stability.

Main Methods:

• Ten able-bodied subjects performed gait-related activities over four measurement days.
• Electromyography (EMG) and kinematics were recorded to assess system performance.
• Three adaptation strategies (entropy, backward prediction, combined) were implemented and compared against baseline and ground truth adaptation.

Main Results:

• Without adaptation, the baseline error rate significantly increased from day 1 to day 2, then plateaued.
• Entropy-based adaptation demonstrated the smallest increase in error rate over the study period.
• Adaptation strategies, particularly entropy-based, showed potential in reducing the impact of concept drift.

Conclusions:

• Concept drift significantly impacts EMG-based control systems, necessitating adaptation strategies.
• Entropy-based adaptation is a simple, feasible, and effective method for reducing error rates in lower limb pattern recognition.
• This study highlights the potential of entropy-based adaptation for enhancing the long-term reliability of myoelectric control systems.