Source-reconstructed EEG graph signal processing: Pitfalls and workarounds
Network Neuroscience · 2026-10-30 · 10 (4) · 967-984
DOI: 10.1162/netn.a.591 ↗Abstract
Abstract Graph signal processing (GSP) enables studying brain structure–function coupling by examining how functional signals vary on the structural connectome (SC). While traditionally used with fMRI, GSP applied to electroencephalography (EEG) is gaining interest due to EEG’s repertoire of brain activities and higher temporal resolution. To this aim, source activities are reconstructed through electrical source imaging (ESI), summarized into parcellated time series via singular value decomposition (SVD), and analyzed using graph Fourier transform based on SC Laplacian’s eigenvectors. This study investigates two methodological biases: SVD polarity ambiguity and ESI spatial leakage. Using simulated epileptic spikes, we validated a method to resolve SVD sign ambiguity and tested it on real interictal epileptogenic discharges. We then quantified the leakage effect on simulation, by comparing graph power spectra between ground truth and its ESI reconstruction, across inverse solutions (exact low-resolution electrical tomography [eLORETA], vs. linear constraint minimal variance) conditions (baseline vs. spike), and connectomes (structural vs. Euclidean distance-based [EC]). We found that uncorrected polarity ambiguity increases high-frequency content, a bias mitigated by the proposed sign control. Conversely, ESI leakage inflates low-frequency content, with stronger effect during spikes, for eLORETA and EC. These results suggest caution in GSP-EEG analyses and inform analytical pipeline selection.
EUVIMED