Assessment of nonnegative matrix factorization algorithms for electroencephalography spectral analysis.

Area of Science:

• Neuroscience
• Signal Processing

Background:

• Nonnegative Matrix Factorization (NMF) is utilized for electroencephalography (EEG) spectral analysis.
• Numerous NMF algorithms exist, but their comparative performance in EEG analysis remains underexplored.

Purpose of the Study:

• To compare the accuracy, stability, and time complexity of four NMF algorithms for EEG data analysis.
• To evaluate the stability of NMF algorithms using a Hierarchical clustering algorithm.

Main Methods:

• Simulated EEG data was used to assess algorithm performance.
• Four NMF algorithms were compared based on estimation accuracy, stability, and time complexity.
• Hierarchical clustering was employed to evaluate NMF stability.

Main Results:

• The hierarchical alternating least squares (HALS) low-rank NMF algorithm (lraNMF_HALS) demonstrated superior performance across all evaluated metrics.
• lraNMF_HALS successfully extracted stable and interpretable features from real resting-state EEG data.

Conclusions:

• The lraNMF_HALS algorithm is recommended for EEG spectral analysis due to its superior accuracy and robustness.
• This study provides a benchmark for selecting NMF algorithms in EEG research.