BCIwiki (bciwiki.com) — A network that explicitly models hemispheric lateralization between frontotemporal electrodes decoded emotions from sparse EEG at up to 87.43% accuracy, researchers at Xiamen University of Technology reported in Cognitive Neurodynamics on September 20, 2026 (PMID 42763501). The Spatiotemporal Spectral Asymmetric Fusion Network (STSANet) uses a dual-branch architecture: one branch captures spectral oscillations while the other quantifies nonlinear activation differences between homologous frontotemporal electrode pairs — a core mechanism of valence regulation. A cross-modal attention mechanism fuses the two feature streams. On the SEED benchmark and a self-collected dataset, STSANet achieved 86.80% and 87.43% accuracy, respectively.
The team, drawn from Xiamen University of Technology, Xiamen Peiyang BCI and Smart Health Innovation Research Institute (Xiamen Peiyang Ruiheng Smart Health Co., Ltd.), and Xiamen Intretech Inc., confirmed that the model remained robust under trial-level cross-validation, and ablation studies identified explicit spatial asymmetry modeling as the key performance driver. The researchers also verified consistent spectral energy distributions between the portable Xmuse headset and the professional-grade Enobio system, suggesting the algorithm could run on consumer-grade EEG hardware. Validation so far covers only two datasets, and cross-population generalization remains untested.