Dynamics-informed priors (DIP) for neural mass modelling

Area of Science:

• Computational neuroscience
• Neuroimaging analysis
• Mathematical modeling

Background:

• Neural mass models (NMMs) are crucial for understanding brain activity.
• Parameter estimation in NMMs, often using dynamic causal modelling (DCM), is vital but sensitive to prior assumptions.
• Limited empirical data and poorly defined priors can bias NMM inference.

Purpose of the Study:

• To develop a computational extension of DCM for improved parameter estimation.
• To establish a strategy for mapping NMM parameters to neuroimaging data.
• To enable data-driven derivation of priors for NMMs in exploratory studies.

Main Methods:

• Proposed DCM with dynamics-informed priors (DIP-DCM), integrating a genetic algorithm (GA) to map parameter values to model dynamics.
• Optimized parameter space sub-regions were identified and translated into parameter priors for DCM.
• DIP-DCM was validated against standard DCM and standalone GA using two neuroimaging datasets.

Main Results:

• DIP-DCM models demonstrated superior predictive accuracy compared to standard DCM and GA.
• The method successfully captured mechanistic signatures of psychiatric conditions and drug effects.
• DIP-DCM effectively navigated local minima and explored parameter spaces guided by model dynamics and data.

Conclusions:

• DIP-DCM offers an advantageous approach to parameter estimation, especially with limited information.
• This method facilitates a data-driven derivation of priors, enhancing exploratory research in neuroscience.
• DIP-DCM shows broad applicability across diverse biological contexts and datasets.