Published September 29, 2026, in Physiological Measurement, this technical study evaluated artifact management for a low-density, dry-electrode eyewear EEG platform.
A selective processing strategy A lightweight gate first identified contaminated windows. Only flagged segments moved to a classifier for eye blinks, horizontal eye movements or facial movements, followed by a denoising network. In 29 participants, the gate reached balanced accuracy of 0.89. Artifact classification reached 0.84, with median macro-F1 of 0.83.
Independent testing An external benchmark drawn from 24 participants produced gate balanced accuracy of 0.82. The end-to-end eyewear evaluation reported median normalized RMSE of 0.19 and spectral cosine similarity of 0.89. Performance varied by participant and artifact type.
What still needs proving The evaluation was offline. Four temporal channels and nose referencing limited spatial information, blinks and eye movements were sometimes confused, and part of the denoising evaluation used semi-synthetic contamination. This does not establish clinical accuracy or embedded consumer readiness.
Source and disclosure Full text checked. The work was conducted in the EssilorLuxottica Smart Eyewear Lab with European Union NextGenerationEU support. Two authors were company employees. Prepared with AI assistance and source-checked September 30, 2026.