Autoregressive Support Sampling for Montage-Agnostic EEG Source Imaging

Abstract

Electroencephalography Source Imaging (ESI) reconstructs cortical activity from scalp measurements. Current methods face three distinct bottlenecks: they require externally computed leadfields, struggle to generalize across varying electrode montages, and often produce deterministic point estimates that ignore spatial uncertainty. We introduce a geometry-conditioned architecture that overcomes these limitations by processing continuous 3D electrode coordinates. The imager predicts an internal leadfield to bypass external forward models at inference, while an autoregressive decoder samples alternative sets of active parcels to model spatial ambiguity. On matched 256-channel synthetic data, our deterministic estimate reduces median localization error by 13.4% to 9.7 mm and increases AUPRC from 0.749 to 0.787. This mapping successfully transfers to real clinical EEG featuring known intracranial stimulation sites. The autoregressive architecture reduces median peak-to-site distance by 15.5% to 22.3 mm relative to the best-performing baseline. Furthermore, training the imager jointly across four standard layouts enables genuine zero-shot transfer to unseen 21-, 76-, and 256-channel montages. On these unseen configurations, the median localization error ranges from 10.4 to 16.5 mm, remaining comparable to the 10.6 mm error achieved on the training layouts.