BCIwiki (bciwiki.com) — A spiking neural network called STAND-Net removes artifacts from EEG signals while consuming 97.98% less power than comparable deep neural networks, researchers reported in IEEE Journal of Biomedical and Health Informatics on March 4, 2026. The authors are Zhang R, Guo X, Pan Y and Gao S.
EEG-based brain-computer interfaces (BCIs) are held back by physiological artifacts that degrade signal fidelity. Deep neural networks (DNNs) improve artifact rejection, but their computational cost has kept them out of wearable BCI systems, the authors wrote.
STAND-Net, short for Spiking Temporal Attention autoeNcoDer Network, relies on event-driven spiking neurons. It pairs a spike-convolution encoder-decoder built on leaky integrate-and-fire neurons, which models spatiotemporal EEG dynamics, with a dilation-enhanced residual backbone for long-range dependencies and a spike-rate attention mechanism that locates artifacts from neuronal firing patterns.
The network improved signal-to-distortion ratio by more than 3.7 dB over state-of-the-art methods across diverse artifacts, and downstream BCI classification accuracy rose 6.64% when using its processed signals, according to the paper. The abstract does not specify the datasets, number of subjects or hardware used.