Towards Efficient Neural Decoder for Dexterous Finger Force Predictions.

Area of Science:

• Robotics
• Neuroscience
• Biomedical Engineering

Background:

• Dexterous robotic hand control necessitates advanced neural-machine interfaces for decoding finger movements.
• Current methods often require extensive multi-finger data for training decoders, leading to high computational demands and large datasets.
• Investigating efficient training strategies for neural decoders is crucial for practical applications.

Purpose of the Study:

• To assess the feasibility of training a neural decoder using limited single-finger surface electromyogram (sEMG) data for predicting multi-finger forces.
• To develop and evaluate a deep forest-based neural decoder for concurrent prediction of three-finger extension and flexion forces.
• To compare the performance of the deep forest decoder against conventional methods.

Main Methods:

• A deep forest model was developed to predict forces for the index, middle, and ring-pinky fingers.
• The model was trained using varying quantities of high-density EMG data under a single-finger training condition.
• Performance was evaluated based on force prediction error and R-squared (R²) values.

Main Results:

• The deep forest decoder achieved a 7.0% force prediction error and an R² of 0.874.
• Performance significantly outperformed conventional EMG amplitude and convolutional neural network (CNN) methods.
• Accuracy decreased with reduced training data and increased noise in testing data.

Conclusions:

• Deep forest decoders demonstrate accurate multi-finger force prediction capabilities.
• The efficiency of deep forest models, characterized by short training times and minimal data requirements, addresses critical needs in neural decoding.
• This research provides valuable insights for developing efficient and accurate neural decoders for advanced robotic hand control and human-machine interaction.