Data-driven mean-field within whole-brain models

Mean-field models provide a link between microscopic neuronal activity and macroscopic brain dynamics. Their derivation depends on simplifying assumptions, such as all-to-all connectivity, limiting their biological realism. Here, we introduce a data-driven framework in which a multi-layer perceptron learns the macroscopic dynamics directly from simulations of a network of spiking neurons. The network connection probability p, inaccessible to analytical treatment, enters the neural network parameterization. Through bifurcation analysis on the trained model, we demonstrate the existence of cusp bifurcation that reshapes the system’s phase diagram in a degenerate manner with synaptic coupling. We integrate this data-driven mean-field into a whole-brain model and show that it performs as well as traditional neural mass models. Lastly, we demonstrate accurate parameter inference for the data-driven mean-field, while the current state-of-the-art models lead to biased estimates. This work presents a flexible and generic framework for building more realistic whole-brain models, bridging the scales. Traditional models of brain activity rely on simplified equations that can be incomplete or inaccurate when biological details are too complex to solve exactly. This work uses machine learning to learn neural dynamics directly from data at the microscale, building a flexible model that matches real brain scans as accurately as established approaches.