An adaptive session-incremental broad learning system for continuous motor imagery EEG classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery electroencephalography (MI-EEG) is crucial for neuro-rehabilitation systems.
• MI-EEG feature spaces evolve with patient recovery, necessitating adaptive recognition models.
• Existing Broad Learning Systems (BLS) struggle with automatic architectural adaptation to complex, time-varying MI-EEG data.

Purpose of the Study:

• To propose an adaptive session-incremental BLS (ASiBLS) for continuous learning in MI-EEG recognition.
• To enable neuro-rehabilitation systems to automatically adapt to evolving patient data.
• To enhance the plasticity and stability of learning models for MI-EEG.

Main Methods:

• Developed an adaptive session-incremental BLS (ASiBLS) integrating mutual information theory.
• Designed a compact temporal-spatial feature extractor (CTS) for initial data.
• Introduced a mutual information maximization constraint (MIMC) for feature distribution alignment in incremental learning (iBLS).

Main Results:

• ASiBLS achieved average decoding accuracies of 79.89% (BCI Competition IV-2a) and 87.04% (BCI Competition IV-2b).
• The model demonstrated effective adaptation across multiple sessions (up to five).
• Evaluations using kappa coefficient and forgetting rate confirmed superior plasticity and stability.

Conclusions:

• ASiBLS successfully addresses the challenge of adapting BLS models to time-varying MI-EEG data.
• The proposed method adaptively generates optimized and reduced models for successive sessions.
• ASiBLS offers improved performance in continuous learning for neuro-rehabilitation applications.