Motor imagery decoding using source optimized transfer learning based on multi-loss fusion CNN

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Transfer learning is vital for decoding complex motor imagery (MI) tasks.
• Existing transfer learning methods often neglect source model optimizability, limiting target domain adaptation and performance.
• This gap hinders effective application in clinical settings like stroke rehabilitation.

Purpose of the Study:

• To develop an optimizable source model for enhanced transfer learning in motor imagery.
• To introduce a novel Source Optimized Transfer Learning (SOTL) framework to improve target domain adaptability.
• To evaluate SOTL's effectiveness in classifying motor imagery tasks for stroke patients.

Main Methods:

• Proposed a Multi-Loss Fusion Convolutional Neural Network (MF-CNN) to create an optimizable source model.
• Developed the SOTL approach to align source model features with target domain characteristics.
• Transferred a model trained on 16 healthy subjects to 16 stroke patients for unilateral upper limb MI tasks.

Main Results:

• Achieved an average classification accuracy of 51.2 ± 0.17% for four types of unilateral upper limb MI tasks in stroke patients.
• Demonstrated significantly higher accuracy compared to traditional deep learning (p < 0.001) and standard transfer learning (p < 0.05).
• Validated the feasibility of using MI models from healthy subjects for stroke patient classification.

Conclusions:

• SOTL significantly enhances the performance of motor imagery decoding in stroke patients.
• The MF-CNN and SOTL framework offer a promising approach for personalized stroke rehabilitation.
• This study provides empirical evidence supporting transfer learning applications in neurological recovery and assistive technology development.