Parameter inference on brain network models with unknown node dynamics and spatial heterogeneity
Model-based data analysis of whole-brain dynamics links the observed data to model parameters in a network of neural masses. In recent years a special focus was placed on the role of regional variance of model parameters for the emergent activity. Such analyses however depend on the properties of the employed neural mass model, which is often obtained through a series of major simplifications or analogies. Here we propose a data-driven approach where the neural mass model needs not to be specified. Building on the recent progresses in identification of dynamical systems with neural networks, we propose a method to infer from the functional data both the neural mass model representing the regional dynamics as well as the region- and subject-specific parameters, while respecting the known network structure. We demonstrate on two synthetic data sets that our method is able to recover the original model parameters, and that the trained generative model produces dynamics resembling the training data both on the regional level and on the whole-brain level. We further apply the method to resting-state fMRI data from Human Connectome Project. We find that to achieve best fit, the model needs two dimensional state space and three regional parameters, one of which is strongly correlated with the map of genetic expression in human brain. The present approach opens a novel way to the analysis of resting-state fMRI with possible applications in understanding the changes of whole-brain dynamics during aging or in neurodegenerative diseases.