MEL Improves EEG-to-fMRI Translation with Coordinate-Preserving Tokenization.

Translating electroencephalography (EEG) data into functional magnetic resonance imaging (fMRI) is a critical challenge in neuroimaging, aiming to infer spatially organized brain activity from fast, accessible electrophysiological recordings. Existing approaches often struggle with the fundamental representation mismatch between EEG and fMRI, which involves differences in temporal lag, channel identity, and frequency-band structure. Researchers propose Multi-band EEG Latent-state Tokenization (MEL), a novel coordinate-preserving EEG representation framework. MEL addresses the mismatch by anchoring each target fMRI response to its preceding EEG history and organizing it into lag-channel-frequency neural-state tokens. This explicit capture of hemodynamic latency and spectral-spatial dynamics allows MEL to align fMRI-relevant EEG representations with capacity-controlled readouts, rather than relying solely on model scaling. Experiments on standard EEG-fMRI benchmarks and external datasets demonstrate that MEL significantly improves prediction accuracy over strong baselines, with ablations confirming that these gains stem from its structured EEG representation.