Beyond the Cortex: A Digital-Twin Simulation Engine for Whole-Brain EEG Source Imaging

EEG provides unparalleled temporal resolution for whole-brain imaging, yet recovering neural sources from scalp recordings is a severely ill-posed problem. Progress toward robust machine-learning solutions is further constrained by the scarcity of ground-truth data, since the true source activity underlying real EEG is never directly observable. This work introduces a high-fidelity simulation engine designed as data-generation infrastructure for AI-based inverse models. While established platforms such as The Virtual Brain (TVB) [1] model large-scale brain dynamics, their EEG simulations primarily focus on neocortical sources. The proposed framework extends TVB by automating the integration of cortical, deep, and cerebellar structures into a unified pipeline, combining high-density cortical parcellations with detailed subcortical atlases and high-resolution structural references, yielding biophysically grounded whole-brain models exceeding 1000 interconnected nodes. Structural connectivity is derived from probabilistic tractography on the Human Connectome Project young-adult dataset [2] via MRtrix3 [3]. To capture spatial heterogeneity, TVB's differentiable stack [4] is extended with node-wise learnable parameters: a structure-specific combination of neural mass models—tailored separately for cortical and subcortical nodes—is fitted jointly against empirical BOLD fMRI and EEG via gradient-based optimization, yielding node-wise parameter maps. Node-wise BOLD fitting improves static functional-connectivity correlation by 36% over a homogeneous baseline, while EEG-driven optimization reproduces the power spectral density of the resting-state phase of a naturalistic simultaneous EEG-fMRI dataset [5]. This same machinery enables individual-specific digital-twin optimization against a subject's empirical recordings. High-resolution tetrahedral meshes with realistic tissue conductivities support DUNEuro finite-element forward modeling [6], projecting neural activity to scalp electrodes with spatiotemporally correlated noise. The framework generates large synthetic EEG datasets with known ground truth and paired BOLD+EEG within a single interpretable model, used to train machine-learning inverse models targeting deep and cerebellar sources traditionally hard to resolve with scalp EEG. Source code and a containerized implementation are publicly released as a reproducible benchmark for data-driven source-reconstruction methods. 1. Ritter, Petra, et al. "The virtual brain integrates computational modeling and multimodal neuroimaging." Brain connectivity 3.2 (2013): 121-145., 10.1089/brain.2012.0120 2. Van Essen, David C., et al. "The Human Connectome Project: a data acquisition perspective." Neuroimage 62.4 (2012): 2222-2231., 10.1016/j.neuroimage.2012.02.018 3. Tournier, J-Donald, et al. "MRtrix3: A fast, flexible and open software framework for medical image processing and visualisation." Neuroimage 202 (2019): 116137., 10.1016/j.neuroimage.2019.116137 4. Pille, Marius, et al. "Fast and easy whole-brain network model parameter estimation with automatic differentiation." bioRxiv (2025): 2025-11., 10.1101/2025.11.18.689003 5. Telesford, Qawi K., et al. "An open-access dataset of naturalistic viewing using simultaneous EEG-fMRI." Scientific Data 10.1 (2023): 554., 10.1038/s41597-023-02458-8 6. Schrader, Sophie, et al. "DUNEuro—A software toolbox for forward modeling in bioelectromagnetism." PloS one 16.6 (2021): e0252431., 10.1371/journal.pone.0252431 Acknowledgements Research supported by the AEGEUS project funded by the European Union, Horizon Europe Programme (GA 101099210) and by the project Ateneo 2025 by Sapienza - University of Rome (code: B83C25004300005).