Enhanced Deep Transfer Learning Model based on Spatial-Temporal driven Scalograms for Precise Decoding of Motor...

Area of Science:

• Biomedical Engineering
• Neuroscience
• Robotics

Background:

• Intelligent rehabilitation robotics (RR) aim to assist post-stroke survivors in regaining limb function.
• Current passive RR systems often limit users to predefined paths, hindering full recovery.
• Decoding patient motion intent is crucial for active, effective robotic rehabilitation.

Purpose of the Study:

• To propose an efficient Transfer Learning based Convolutional Neural Network (TL-CNN) model for decoding post-stroke patients' motion intentions.
• To introduce Spatial-Temporal Descriptor based Continuous Wavelet Transform (STD-CWT) as an input for TL-CNN to optimize limb movement intent decoding.
• To enable dexterously active robotic training for improved functional recovery in stroke survivors.

Main Methods:

• Developed a Transfer Learning based Convolutional Neural Network (TL-CNN) model.
• Utilized Spatial-Temporal Descriptor based Continuous Wavelet Transform (STD-CWT) with Morse, Amor, and Bump wavelets as input for the TL-CNN.
• Compared STD-CWT with the standard Continuous Wavelet Transform (CWT) using electromyogram signals from five stroke survivors performing 21 motor tasks.

Main Results:

• The proposed STD-CWT method achieved significantly higher decoding accuracy (p<0.05) and faster convergence compared to the standard CWT.
• Demonstrated clear class separability for individual motor tasks across different subjects.
• STD-CWT Scalograms showed robust decoding of motor intention patterns.

Conclusions:

• The STD-CWT method facilitates robust decoding of motor intention for stroke rehabilitation.
• This approach can enable intuitive and active motor training in rehabilitation robotics (RR).
• The findings support the use of STD-CWT Scalograms for precise, multi-class motor task decoding, paving the way for full motor function restoration through active RR.