Multiparameter optimization for whole-brain personalization

Advancing therapies for neurological disorders relies on developing efficient, personalized whole-brain models that can capture mechanistic insights and pathological features. Current challenges in creating such computational models arise from managing a considerable number of adjustable parameters. The conventional manual exploration of this parameter space is a serious limitation for personalizing models to individual subject data [1]. This work aims to address this by approaching the problem as an optimization challenge in the context of developing brain models capable of characterizing individual pathology, particularly Alzheimer’s disease (AD), with the ultimate goal of providing better transcranial electromagnetic stimulation protocols. For this purpose, we work with hybrid brain models [1], combining biophysical head models and a network of laminar neural mass models (LaNMM), each capable of representing slow (alpha) and fast (gamma) activity [2]. For the personalization of whole-brain models, our approach includes exploring various parameters that can be tuned to better fit the pathology and the data, which can be either fMRI or a combination of fMRI and EEG. After an initial inspection of the dynamical landscape, we identify a parameter space that reflects realistic brain dynamics. Subsequently, we define an apt similarity function between the simulated and real data. To maximize this function and achieve optimal parameter estimation, we employ Bayesian Optimization. This method, extensively used in Machine Learning, offers a good compromise between computational time and goodness of fit, proving to be effective for our models. Finally, we present some initial results on a small set of personalized whole-brain models for both healthy and AD patients.