A Semi-Supervised Progressive Learning Algorithm for Brain-Computer Interface.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Brain-computer interfaces (BCI) often face challenges with low recognition accuracy and lengthy calibration periods.
• Identifying motor imagery tasks and classifying fine-grained motions using electroencephalogram (EEG) and electroencephalogram-electromyogram (EEG-EMG) fusion analysis are particularly difficult.

Purpose of the Study:

• To develop an end-to-end semi-supervised learning framework to address the limitations in BCI accuracy and calibration time.
• To improve EEG classification and EEG-EMG fusion analysis for enhanced BCI performance.

Main Methods:

• Implemented a metric learning-based label estimation strategy, sampling criterion, and progressive learning scheme for efficient feature extraction from unlabeled EEG data.
• Utilized synchronous EMG features as pseudo-labels for unlabeled EEG samples to extract deep-level features from synergistic EEG-EMG complementarity.
• Employed deep encoders for feature extraction and performance enhancement in hybrid BCI systems.

Main Results:

• Achieved a 5.40% improvement on BCI Competition IV Dataset IIa using 80% unlabeled samples and an average 3.35% improvement on two public BCI datasets.
• Demonstrated a 5.53% improvement for the Upper Limb Motion Dataset and an average 4.34% improvement on two hybrid datasets by incorporating EMG features.
• Ablation experiments confirmed substantial performance improvements in deep encoders, averaging 5.53%.

Conclusions:

• The proposed framework significantly enhances the performance of deep networks in BCI systems.
• The framework substantially reduces calibration time for EEG-EMG fusion analysis.
• This approach holds great potential for developing efficient, high-performance hybrid BCIs for motor rehabilitation.