Motor imagery classification via combinatory decomposition of ERP and ERSP using sparse nonnegative matrix...

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Brain activity analysis utilizes multimodal data (EEG, MEG, MRI) across time, frequency, and space.
• Nonnegative Matrix Factorization (NMF) excels in pattern extraction but is limited by negative brain signal values.
• Existing NMF studies for brain-computer interfaces primarily use spectral features, neglecting time-domain information and signal sparsity.

Purpose of the Study:

• Develop a novel NMF-based method for motor imagery classification.
• Integrate time-domain (ERP) and frequency-domain (ERSP) brain signal features.
• Incorporate sparsity constraints into NMF for improved analysis of electric brain signals.

Main Methods:

• Introduced a modified mixed alternating least squares-based NMF (MALS-NMF).
• Combined event-related potential (ERP) and event-related spectral perturbation (ERSP) features.
• Applied sparsity constraints to coefficient and basis matrices during factorization.

Main Results:

• The MALS-NMF method demonstrated effectiveness in motor imagery classification.
• Imposing sparsity constraints on ERP and ERSP factorization components enhanced performance.
• Extensive experiments confirmed the superiority of the proposed method over eight other approaches.

Conclusions:

• MALS-NMF provides an effective solution for motor imagery classification.
• The method offers new insights into brain dynamics pattern analysis.
• This approach advances the application of NMF in brain-computer interfaces.