Spectral Independent Component Analysis with noise modeling for M/EEG source separation.

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Independent Component Analysis (ICA) is widely used for electroencephalography (EEG) and magnetoencephalography (MEG) signal processing.
• Traditional ICA methods assume noiseless data and non-Gaussian sources, limiting their application.
• Existing methods often require a preliminary Principal Component Analysis (PCA) step, which can be detrimental.

Purpose of the Study:

• Introduce the Spectral Matching ICA (SMICA) model for analyzing EEG and MEG signals.
• Address limitations of traditional ICA by incorporating noise and Gaussian sources.
• Improve the accuracy and efficiency of source separation in neuroimaging data.

Main Methods:

• Model signals as a linear mixture of independent Gaussian sources with additive noise.
• Utilize the Gaussian assumption to simplify the negative log-likelihood as a sum of spectral covariance matrix divergences.
• Employ the Expectation-Maximization (EM) algorithm for parameter estimation.

Main Results:

• SMICA achieved a median dipole localization error of 1.5 mm on phantom MEG data, significantly outperforming competing methods (≥7 mm).
• On EEG datasets, SMICA identified a source subspace with less pairwise mutual information and better dipolar characteristics.
• SMICA identified over 80% of strongly dipolar sources (dipolarity >90%) with 10 sources, compared to <65% for competing methods.

Conclusions:

• SMICA provides a robust alternative to traditional noiseless ICA models.
• The number of sources in SMICA is controlled by the mixing matrix size, avoiding detrimental dimension reduction steps.
• SMICA demonstrates superior performance in dipole localization and source separation for EEG/MEG data.