Mitigating activity mixing with personalized whole-brain modeling
Mechanistic neuroimaging-based biomarkers based on localized brain activities or interactions are a central tenet in precision and personalized psychiatry. However, their accuracy may be limited by Activity Mixing that is a novel construct indicating the entanglement of any local neuronal activity with activities elsewhere in the network through long-range spatiotemporal correlations, which degrades the localization of brain-symptom associations. Here, we posit that the ill-posed inverse problem of activity mixing can be mitigated by fitting of generative whole-brain models. We developed a multi-objective fitting approach to estimate subject-specific local- and inter-areal brain-dynamics control parameters from neuroimaging observables. By integrating both synchronization and criticality metrics, this approach yields personalized parameters capture individual brain network dynamics more accurately than raw observables. In silico validation demonstrated that fitted model parameters improved brain-symptom correlation estimates by 30-85% and reduced false-negative rates by approximately [~]67% relative to conventional observables-based analyses. As in vivo proof-of-concept, resting-state magnetoencephalography (MEG) data from 230 patients with major depressive disorder (MDD) showed that aberrant brain criticality in the alpha-frequency band (11 Hz) was a significant predictor of disability with a correlation coefficient of 0.236 (95% confidence interval (CI) = [0.206, 0.266]) in 27 (CI = [23, 31]) significant cortical parcels. Model fitting both improved this correlation estimate by [~]56% up to 0.368 (CI = [0.341, 0.405]) and localized it [~]25% more narrowly to 20 (CI = [18, 22]) parcels. These findings suggest that model fitting can mitigate the effects of activity mixing and provide control-parameter estimates that delineate mechanistic biomarkers for brain disorders more accurately than the raw brain imaging observables.