Solving large-scale MEG/EEG source localisation and functional connectivity problems simultaneously using state-space...

Area of Science:

• Neuroscience
• Computational Biology
• Statistical Modeling

Background:

• State-space models are crucial for understanding unobserved dynamics in diverse fields.
• Conventional methods like Kalman filtering face computational challenges in large-scale analyses.
• Sparse estimators introduce bias-variance trade-offs, especially in low signal-to-noise ratio (SNR) scenarios.

Purpose of the Study:

• To propose Multiple Penalised State-Space (MPSS) models for efficient analysis of complex dynamics.
• To develop novel algorithms for solving MPSS models using backpropagation, gradient descent, and alternating least squares.
• To introduce K-fold cross-validation for robust regularisation parameter selection.

Main Methods:

• Introduction of data-driven regularisation within MPSS models.
• Development of novel optimization algorithms (backpropagation, gradient descent, alternating least squares).
• Application of K-fold cross-validation for parameter tuning.

Main Results:

• MPSS models demonstrate effective performance in simulations across various SNR conditions.
• Successful application to large-scale synthetic magneto- and electro-encephalography (MEG/EEG) data.
• Accurate concurrent brain source localization and functional connectivity analysis on real MEG/EEG data.

Conclusions:

• The MPSS framework overcomes computational and scalability limitations of existing state-space models.
• Enables detailed analysis of cognitive brain functions on large-scale neuroimaging data.
• Facilitates accurate exploration of brain source localization and functional connectivity.