Person thresholds from 60 seconds per state increased one-month eye-state balanced accuracy by 7.6 percentage points in 30 ds004148 adults using offline 8-13 Hz EEG spectra

We tested first-visit person calibration for recorded eye-state classification one month later. We used offline posterior 8-13 Hz spectra from OpenNeuro ds004148 version 1.0.1. A frozen split assigned 30 adults to population training and 30 to evaluation. The population threshold minimized balanced error on first-visit windows. The person threshold used 60 seconds per state from the first visit. Mean balanced accuracy was 73.15% with the population threshold and 80.77% with the person threshold. The paired gain was 7.62 percentage points. Its 95% person bootstrap interval was 4.08 to 12.76 points. Each of 10,000 draws refit the population threshold. The person rule improved 21 people, harmed five, and tied in four. No person was excluded. The package preserves all 360 input spectra, predictions, source hashes, code, and reproduction steps. Independent calculations, four fresh raw replays, a fresh metadata check, and a full repeat passed. The result applies to these offline files and fixed rules. State order, prior eye-artifact removal, and full-record processing limit broader claims.

electroencephalography · alpha power · eye-state classification · person calibration · balanced accuracy · test-retest · EEG spectra · threshold classification