Stored PSG AI labels had 59% mean recall of human stage changes within 60 seconds on common valid epochs in 100 BOAS people

We tested rate and event agreement in OpenNeuro ds005555 v1.1.3. The stored labels cover 128 nights from 100 people. The main analysis used 114,040 consecutive edges with valid human, headband AI, and PSG AI labels at both ends. Each person had equal weight. PSG AI had 9.96 stage changes per eligible hour. The human reference had 10.07. Their rate difference interval included zero and did not establish equivalence. PSG AI had 59.20% mean recall of human changes in a maximum one-to-one match within 60 seconds. Its 95% person bootstrap interval was 56.48% to 61.85%. Recall was 44.05% when the ordered stage pair also had to agree. PSG epoch agreement was 86.89%. Headband AI had 51.51% recall at 60 seconds. The locally planned primary headband rate difference was -1.78 changes per hour, with an interval from -2.36 to -1.20. On all human-valid edges, exact missing-label bounds allowed a mean difference from -2.12 to +2.33 changes per hour. Thus the deficit applies to common valid periods. It is not established for the full human-valid record. The human consensus is a reference, not physiological truth. A related publication appeared during this run and reported the headband result. This study adds the planned PSG event tests on the same cohort. The package preserves source tables, fixed URLs and hashes, analysis code, event ledgers, bootstrap draws, proofs, and reproduction instructions. Independent algorithms checked counts, matches, bounds, and all bootstrap means. A fresh source download reproduced all ten numerical files byte for byte.

sleep staging · sleep stage transitions · event agreement · recall · precision · F1 score · polysomnography · wearable EEG · human labels · artificial intelligence · BOAS