Bayesian fusion and multimodal DCM for EEG and fMRI

Area of Science:

• Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• Characterizing functional brain architectures often relies on single neuroimaging modalities like EEG or fMRI.
• Each modality offers complementary strengths: EEG provides high temporal resolution, while fMRI offers high spatial resolution.
• Integrating these modalities could yield a more complete picture of brain dynamics.

Purpose of the Study:

• To evaluate if integrating multimodal EEG and fMRI data provides superior characterization of functional brain architectures compared to using either modality alone.
• To introduce and demonstrate a Bayesian fusion method for combining EEG and fMRI data within a dynamic causal modeling (DCM) framework.

Main Methods:

• Developed a dynamic causal model capable of generating both EEG and fMRI data from shared neuronal dynamics.
• Introduced Bayesian fusion to incorporate empirical neuronal priors from EEG-based DCM into fMRI-based DCM.
• Generated synthetic EEG and fMRI time-series data for an auditory oddball paradigm using biologically plausible parameters derived from empirical EEG data.

Main Results:

• Bayesian fusion significantly improved model evidence (marginal likelihood) compared to inverting fMRI data alone, indicating more efficient parameter estimation.
• Quantified information gain using Kullback-Leibler divergence, showing EEG data improved estimates of hemodynamic parameters.
• Demonstrated that Bayesian fusion is necessary to resolve conditional dependencies between neuronal and hemodynamic estimators.

Conclusions:

• Bayesian fusion of EEG and fMRI data offers a powerful approach to exploit the complementary temporal and spatial resolutions of each modality.
• This multimodal integration provides a more accurate and comprehensive characterization of functional brain architectures.
• The proposed method is applicable to any multimodal dataset explainable by a DCM with common neuronal parameterization.