Towards an efficient validation of dynamical whole-brain models.

Area of Science:

• Computational neuroscience
• Neuroimaging analysis
• Mathematical modeling

Background:

• Accurate simulation of resting-state brain dynamics requires optimal parameter selection in whole-brain models.
• High-dimensional parameter spaces make traditional grid search (GS) optimization computationally infeasible.

Purpose of the Study:

• To evaluate alternative optimization methods for whole-brain models.
• To compare the performance and efficiency of Nelder-Mead Algorithm (NMA), Particle Swarm Optimization (PSO), Covariance Matrix Adaptation Evolution Strategy (CMAES), and Bayesian Optimization (BO) against GS.

Main Methods:

• Utilized an ensemble of coupled phase oscillators based on empirical structural connectivity from 105 healthy subjects.
• Assessed optimization performance in two- and three-dimensional parameter spaces.
• Used dense grid search as a benchmark for comparison.

Main Results:

• Alternative methods (NMA, PSO, CMAES, BO) demonstrated competitive fitting quality compared to GS.
• CMAES and BO showed marked differences in computational resources and stability.
• For 3D parameter spaces, CMAES and BO achieved similar results to GS in less than 6% of the computation time.

Conclusions:

• CMAES and BO are efficient and viable alternatives to high-dimensional GS for whole-brain model parameter optimization.
• These findings facilitate efficient validation of models for personalized brain dynamics simulations.