Kalman-filter-based EEG source localization

Abstract

This thesis uses the Kalman filter (KF) to solve the electroencephalographic (EEG) inverse problem to image its neuronal sources. Chapter 1 introduces EEG source localization and the KF and discusses how it can solve the inverse problem. Chapter 2 introduces an EEG inverse solution using a spatially whitened KF (SWKF) to reduce the computational burden. Likelihood maximization is used to fit spatially uniform neural model parameters to simulated and clinical EEGs. The SWKF accurately reconstructs source dynamics. Filter performance is analyzed by computing the innovations’ statistical properties and identifying spatial variations in performance that could be improved by use of spatially varying parameters. Chapter 3 investigates the SWKF via one-dimensional (1D) simulations. Motivated by Chapter 2, two model parameters are given Gaussian spatial profiles to better reflect brain dynamics. Constrained optimization ensures estimated parameters have clear biophysical interpretations. Inverse solutions are also computed using the optimal linear KF. Both filters produce accurate state estimates. Spatially varying parameters are correctly identified from datasets with transient dynamics, but estimates for driven datasets are degraded by the unmodeled drive term. Chapter 4 treats the whole-brain EEG inverse problem and applies features of the 1D simulations to the SWKF of Chapter 2. Spatially varying parameters are used to model spatial variation of the alpha rhythm. The simulated EEG here exhibits wave-like patterns and spatially varying dynamics. As in Chapter 3, optimization constrains model parameters to appropriate ranges. State estimation is again reliable for simulated and clinical EEG, although spatially varying parameters do not improve accuracy and parameter estimation is unreliable, with wave velocity underestimated. Contributing factors are identified and approaches to overcome them are discussed. Chapter 5 summarizes the main findings and outlines future work.