In clinical neurology, demographic changes and increases in diagnostic complexity coupled with limited financial resources call for novel AI-driven yet mathematically interpretable technologies such as digital twins. Digital twins of the brain are usually models based on differential equations representing large-scale brain network dynamics and interactions (i.e., neural mass model). One main barrier for translating digital twin technologies based on neural mass models to clinical translation is the lack of uncertainty quantification and high computational costs. As a potential solution to bridge the barrier is to go beyond classical parameter inference through grid search, Bayesian optimization such as simulation-based inference offers a practical solution for efficient parameter interference at low computational cost and providing a probabilistic mapping instead of point-estimates. Simulation-based inference constitutes a novel combination of deep neural networks and Bayesian inference to learn the mapping of parameters to the model output and can be easily transferred to experimental data. At different levels of model complexity (i.e., number of parameters) I will showcase the added value of simulation-based inference in comparison to classical inference methods, making it most suitable for real-life parameter inference of large scale biomedical datasets, coupled a suggestion of the optimal level of model complexity that represents neural dynamics best through Bayesian model comparison based on deep neural networks. I will finish with a formulation of needs for further method development from the clinical side.