← Projects · research

EEG microstate representation learning

Compares image-based scalp maps with maps built from connections between recording electrodes. When both use the same processing step on 203 resting-state recordings, the electrode-based approach accounts for 61.8% of the signal versus 54.2% for the image-based approach. The image-based approach produces tighter groups of activity.

This project studies discrete EEG microstate representation learning with Conv-VaDE and Graph-VaDE. Conv-VaDE learns from scalp topographic maps, while Graph-VaDE represents electrodes as graphs. Both use learnable Gaussian-mixture priors for probabilistic segmentation.

On synthetic data with known ground truth, Graph-VaDE retained an adjusted Rand index of 1.0 and global explained variance of 0.85. The source code is publicly available.