Neural network committees for finger joint angle estimation from surface EMG signals

Area of Science:

• Biomedical Engineering
• Neuroscience
• Virtual Reality

Background:

• Virtual reality (VR) systems typically use hand and finger positioning for control.
• Surface electromyography (SEMG) offers a potential for more intuitive and unconstraining biocontrol of VR environments.

Purpose of the Study:

• To develop a technique for predicting finger joint angles from SEMG signals using neural networks.
• To enable direct biocontrol of VR environments through muscle electrical activity.

Main Methods:

• Collected SEMG and actual finger joint angle data during index finger flexion-extension at varying speeds.
• Trained multiple neural networks to predict joint angles from SEMG parameters.
• Formed committees of the best-performing neural networks for evaluation.

Main Results:

• Neural network committees accurately predicted finger joint angles despite hysteresis in SEMG signals.
• Root-mean-square errors for prediction ranged from 0.085 to 0.163 degrees.
• Prediction accuracy was consistent across different speeds of finger movement.

Conclusions:

• Neural network committees effectively predict finger joint angles from SEMG signals.
• This technique holds promise for enhancing user interaction and control in virtual reality.