Online decoding of rat self-paced locomotion speed from EEG using recurrent neural networks

A research team has achieved a breakthrough in non-invasive neural decoding, accurately predicting rats' self-selected running speed directly from skull-surface EEG signals. The system, developed by Alejandro de Miguel and colleagues, uses recurrent neural networks (RNNs) to process streaming EEG data from 32 electrodes, achieving an impressive 0.88 correlation (R²=0.78) between predicted and actual locomotion speed on a non-motorized treadmill.

Technically, the approach represents a significant advance over previous methods that required invasive implants or achieved only modest accuracy. The team trained their models on an extensive dataset of over 133 hours of recordings, finding that low-frequency oscillations (<8 Hz) from visual cortex electrodes provided the strongest predictive signals. Notably, the system demonstrated session-to-session generalization within individual animals, though cross-animal transfer remained challenging.

The research reveals that cortical states carry information not just about current movement but also about future and past dynamics extending up to 1000 milliseconds. This temporal encoding suggests richer neural representations of action than previously understood. The work establishes a framework for developing high-performing, non-invasive BCI systems that could eventually translate to human applications in rehabilitation and prosthetic control, while advancing fundamental neuroscience of distributed action representations.