Reservoir computing reconstructs blood-oxygen-level-dependent signals: whole-brain modeling study

European Physical Journal: Special Topics. 2025. P. 1–8.

Badarin A., Klinshov V., Smelov P., Andreev A.

Understanding and reconstructing brain dynamics from partial or noisy neuroimaging data remains a critical challenge in computational neuroscience. This study presents a novel framework combining neural mass modeling and reservoir computing (RC) to recover missing blood-oxygen-level-dependent (BOLD) signals while preserving functional connectivity patterns. We first simulate whole-brain dynamics using a Wilson–Cowan neural mass model with biologically realistic structural connectivity, optimizing parameters to match empirical functional connectivity matrices. Next, we employ RC to reconstruct individual BOLD signals using only the remaining signals as inputs. Our results demonstrate that RC achieves high-fidelity signal recovery, particularly for strongly interconnected regions. Crucially, the functional connectivity matrices derived from reconstructed signals show near-perfect agreement with the original simulated matrices, despite minor amplitude discrepancies. This work establishes RC as an effective tool for neuroimaging data reconstruction, with direct applications in both research and clinical settings where data loss or artifacts may compromise analyses.

Keywords: BOLD signal, Neural mass model, Reservoir computing