BCIwiki (bciwiki.com) — Sequences of EEG microstates can serve as triggers for a brain-computer interface (BCI), with transfer learning pushing left- versus right-hand motor imagery classification to peak accuracies around 89%, researchers at OTH Regensburg reported in a preprint posted to bioRxiv on August 23, 2026. The work has not been peer reviewed.
EEG microstates are a small number of quasi-stable spatial patterns of brain activity, and their trajectories over time are thought to reflect neural processing. The team used a semi-supervised deep learning architecture, an LSTM-based autoencoder paired with a dense neural network, to classify imagined left- and right-hand movements from these sequences, using the classifier output as the BCI trigger.
The authors compared a two-stage approach, in which the autoencoder and classifier were trained separately, with an end-to-end approach that combined reconstruction and classification losses. According to the abstract, the model extracted relevant features for within-subject and within-session classification, and transfer across sessions or across subjects produced peak accuracies around 89%.
Using the amount of transfer-learning data required as a proxy for calibration time, the researchers found that about 400s of calibration were needed on average to reach 80% accuracy. They say microstate analysis reduces multichannel EEG to a limited set of brain states over time and could support more user-friendly, calibration-free BCIs, a goal the study has not yet demonstrated.