BCIwiki (bciwiki.com) — A neural network called MTANet decoded motor imagery EEG signals with up to 89.47% average accuracy on a public benchmark, outperforming seven existing models, researchers at Chongqing University reported in Behavioural Brain Research on September 20, 2026 (PMID 42764074).
MTANet uses two core modules: an EEG inception attention module that extracts and weights temporal-frequency-spatial features from raw EEG, and a parallel temporal sequence attention module that processes weighted segments of different feature sequences through parallel temporal convolutional networks. The architecture is end-to-end, requiring no manual feature engineering.
On three public motor imagery datasets, MTANet achieved average accuracies of 83.39% (plus or minus 0.09) on BCI Competition IV dataset IIa, 89.47% (plus or minus 0.08) on dataset IIb, and 77.06% (plus or minus 0.12) on BCI Competition III dataset IIIa. The team is based at the State Key Laboratory of Power Transmission Equipment and System Security and New Technology at Chongqing University's School of Electrical Engineering. The study was validated offline on competition datasets and has not been tested in an online BCI setting.