Data-driven fusion of EEG, functional and structural MRI: A comparison of two models

It has become quite common for multiple brain imaging types to be collected for a particular study. This raises the issue of how to combine these imaging types to gain the most useful information for inference. One can perform data integration, where one modality is used to improve the results of another, or true data fusion, where multiple modalities are used to inform one another. We propose two new methods of data fusion, entropy bound minimization (EBM) for joint independent component analysis (jICA) and independent vector analysis with a Gaussian prior (IVA-G), and compare them to the established data fusion techniques of multiset canonical correlation analysis (MCCA) and jICA using Infomax. Additionally, we propose a simulation model and use it to probe the limitations of these methods. Results show that EBM with jICA outperforms the other selected methods but is highly sensitive to the availability of joint information provided by these modalities.