Decomposition of neurological multivariate time series by state space modelling

Area of Science:

• Neuroscience
• Statistical Time Series Analysis

Background:

• Multivariate time series decomposition is crucial for neuroscience data analysis.
• Existing methods like Factor Analysis (FA) and Independent Component Analysis (ICA) have limitations in capturing temporal dynamics.

Purpose of the Study:

• To introduce linear state space modeling as a superior method for time series decomposition in neuroscience.
• To demonstrate its capability in separating components with diverse spectral properties and identifying artifacts.

Main Methods:

• Review of FA and ICA for time series decomposition.
• Application of linear state space modeling, a generalization of ARMA modeling.
• Exploitation of temporal dynamics for component separation.
• Discussion of generalizations including relaxed independence, non-stationary noise, and non-Gaussianity.

Main Results:

• State space modeling successfully decomposes time series, capturing dynamical information missed by FA and ICA.
• Separation of components with sharp (e.g., alpha, sleep spindles) and broad (e.g., fMRI artifacts, epileptic spikes) power spectra is achieved.
• Demonstrated effectiveness in analyzing electrocardigram and electroencephalogram (EEG) data, including artifact removal and identification of pathological components in epilepsy patients.

Conclusions:

• Linear state space modeling provides a more detailed and powerful approach to time series decomposition in neuroscience.
• This method enhances the analysis of complex biological signals, particularly in identifying and removing artifacts and characterizing neural activity.