Using Machine Learning to Fit Whole Brain Models to Concurrent Resting-State EEG and fMRI Data

Abstract

Whole Brain Modelling (WBM) uses mesoscale computational models of neural activity to study neuroimaging phenomena and medical interventions. Often WBM studies involve selecting a computational model, fitting the model to some neuroimaging data, and then performing experiments on the fitted model. The focus of this thesis is the model fitting aspect, optimizing multimodal WBMs to reproduce phenomena from concurrent resting-state electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) recordings. The goal is to effectively utilize the high temporal resolution of EEG combined with the high spatial resolution of fMRI to generate biologically meaningful parameter sets. The deep learning methodology used implements WBMs in PyTorch as custom Continuous Time Recurrent Neural Networks (CTRNN), which can be optimized by backpropagation through time (BPTT). This approach has been explored for fitting parameters of the Reduced Wong Wang Excitatory Inhibitory (RWWEI) model to resting-state fMRI (Griffiths & Wang et al., 2022) and fitting parameters of the Jansen-Rit model to evoked potential EEG (Momi et al., 2023). It is different from regular ML approaches, in that each parameter fit has a predefined interpretation that will not change, which adds challenges to finding local optimum regimes, but retains biological interpretability. It has the potential to fit significantly more parameters simultaneously, compared to other approaches such as brute force search. The performance of fitting to resting-state EEG with a power spectral density (PSD) objective function, as well as fitting to true-time scale blood oxygen level dependent (BOLD) signal from fMRI with a novel fitting paradigm is tested. Subsequently, fitting both modalities together in the same objective function is evaluated. Working with a multimodal model opens the possibility to study the phase relationship between synthetic resting-state EEG and fMRI. This motivates and enables performing a dynamical systems analysis of the alpha-BOLD anticorrelation phenomenon, which relates the phase of empirical EEG and BOLD time series. A case study is presented using the RWWEI model to demonstrate one approach for which models can be used in research once the parameters have been fit. Additionally, the analysis establishes ground truth behavior of the model which can be used to evaluate parameter fitting performance.