This study tackles the challenge of accurately inverting neural and vasodilatory signals from BOLD-fMRI data. We propose a novel framework that substantially enhances prediction performance through: • Multi-scale dynamic feature extraction to comprehensively characterize BOLD signal properties. • A stacking ensemble architecture that synergistically combines multiple heterogeneous base learners. • Hierarchical model fusion via a meta-learner to robustly integrate predictions and capture complex nonlinear mappings. Evaluated on synthetic data from the Balloon model, our method achieves an R² of 0.92 for the vasodilatory signal and 0.78 for the neural drive signal, outperforming existing benchmarks.Validation on real fMRI data shows successful reconstruction of neural activity, with reconstructed BOLD signals correlating with measured signals at levels up to 0.9931.. This provides a new pathway for high-fidelity inversion of microscopic neural activity.
Yan et al. (Mon,) studied this question.