Generative modeling offers a practical way to obtain labeled clinical electroencephalography (EEG) data under strict privacy constraints. Recent generative models report large gains in fidelity, yet these gains hide a fundamental flaw. EEG is not an arbitrary multichannel signal: each channel records the voltage difference between two scalp electrodes, so real recordings must satisfy exact linear identities across channels. Current generators routinely violate these identities without being noticed.
We therefore introduce MontageFlow, a flow‑matching model that enforces physical validity by restricting its output to signals that satisfy the linear identities. Ensuring validity also improves fidelity. To preserve EEG characteristics that oversmoothing erases, we condition the model on a physiological state vector encoded from the whole EEG window; this vector is sampled from a prior fitted in latent space.
Across three large‑scale benchmarks, MontageFlow improves TS‑FID by 1.6‑3.6× over the closest rival while guaranteeing validity.
Review