A local and global collaborative optimization method for brain network models

Data-driven brain network models (BNM) view the brain as a complex network system, which can simulate the brain dynamics and reveal the neural mechanisms of brain diseases. However, existing modeling methods rarely consider the heterogeneity of local parameters, and primarily focus on the differences in connection characteristics between empirical and simulated neural activity, thus overlooking the global characteristics of neural activity. To address the above issues, we proposed a local and global collaborative optimization method for BNM. The neural mass model (NMM) is applied to simulate the dynamics of local brain regions, and the phase transfer entropy (PTE) network is employed to couple different brain regions. The local parameters are estimated by the improved chimp optimization algorithm and the Riemann geometry. The global loss function is defined as the average symmetric Kullback-Leibler (KL) distance between the empirical and simulated electroencephalogram (EEG) microstate features, including duration and occurrence. The optimal fitted model is obtained by adjusting the global coupling coefficient. The estimated local parameters reveal the neural mechanisms underlying the reduced consciousness level in patients with disorders of consciousness (DOC), characterized by increased inhibitory activity in the parietal and occipital lobes, while decreased excitatory activity in the frontal and parietal lobes. The simulated EEG of constructed BNM can reproduce the microstate duration and occurrence of the empirical EEG, and thus accurately characterize the global neural activity pattern of empirical EEG. We also demonstrated that using directed networks to couple different NMMs and performing local parameter optimization can improve model fitting accuracy.