Multi-scale brain simulation with integrated positron emission tomography yields hidden local field potential activity that augments machine learning classification of Alzheimer’s disease

While the prevalence of neurodegenerative diseases and dementia increases, our knowledge of the underlying pathomechanisms and related diagnostic biomarkers, outcome predictors, or therapeutic targets remains limited. In this article, we show how computational multi-scale brain network modeling using The Virtual Brain (TVB) simulation platform supports revealing potential disease mechanisms and can lead to improved diagnostics.TVB allows standardized large-scale structural connectivity (SC)-based modeling and simulation of whole-brain dynamics. We combine TVB with a cause-and-effect model for amyloid-beta, and machine-learning classification with support vector machines and random forests. The amyloid-beta burden as quantified from individual AV-45 PET scans informs parameters of local excitation/inhibition balance. We use magnetic resonance imaging (MRI), positron emission tomography (PET, specifically Amyloid-beta (Abeta) binding tracer AV-45-PET, and Tau-protein (Tau) binding AV-1451-PET) from 33 participants of Alzheimer’s Disease Neuroimaging Initiative study 3 (ADNI3). The frequency compositions of simulated local field potentials (LFP) are under investigation for their potential to classify individuals between Alzheimer’s disease (AD), Mild Cognitive Impairment (MCI), and healthy controls (HC) using support vector machines and random forest classifiers.The combination of empirical features (subcortical volumetry, AV-45- and AV-1451-PET standard uptake value ratio, SUVR per region) and simulated features (mean LFP frequency per brain region) significantly outperformed the classification accuracy of empirical data alone by about 10% in the accuracy index of weighted F1-score (empirical 64.34% vs. combined 74.28%). There was no significant difference between empirical and simulated features alone. The features with the highest feature importance showed high biological plausibility with respect to the AD-typical spatial distribution of the features. This was demonstrated for all feature types, e.g., increased importance indices for the left entorhinal cortex as the most important Tau-feature, the left dorsal temporopolar cortex for Abeta, the right thalamus for LFP frequency, and the right putamen for volume.In summary, here we suggest a strategy and provide proof of concept for TVB-inferred mechanistic biomarkers that are direct indicators of pathogenic processes in neurodegenerative disease. We show how the cause-and-effect implementation of local hyperexcitation caused by Abeta can improve the machine-learning-driven classification of AD. This proves TVBs ability to decode information in empirical data by means of SC-based brain simulation.