Conv-VaDE makes EEG microstate discovery interpretable with deep learning

A new paper from researchers at Saheed Faremi, Andrea Visentin, and Luca Longo presents Conv-VaDE (Convolutional Variational Deep Embedding), a deep learning approach to EEG microstate analysis that replaces traditional hard clustering with probabilistic soft assignment. Unlike conventional methods like Modified K-Means, Conv-VaDE jointly learns a latent representation and a generative decoder, allowing it to reconstruct interpretable scalp topographies from cluster prototypes. This enables researchers to verify what each discovered brain state actually looks like on the scalp.

The model was evaluated on the LEMON resting-state eyes-closed EEG dataset with ten participants. The team conducted a systematic architecture search across cluster count (K=3 to 20), latent dimensionality, network depth, and channel width. The best-performing configuration used depth L=4, achieving a global explained variance (GEV) of 0.730 and a silhouette score of 0.229 at K=4. The search showed that moderately deep networks with compact channel widths and small latent dimensions consistently outperformed larger models, demonstrating that careful architecture design—not scale—is key to interpretable EEG microstate discovery.

This work addresses a critical gap: traditional methods offer no learned latent representations and no way to decode latent configurations into verifiable scalp topographies, limiting transparency. Conv-VaDE's polarity invariance scheme and probabilistic assignments bring new levels of interpretability to brain state analysis, with potential applications in cognitive neuroscience, clinical diagnostics, and brain-computer interfaces.