Ten-trial calibration had a theta memory error prediction gain interval that included zero in 13 SEEG patients at equal training size under independent patient sampling

We tested whether ten early labeled trials improved later memory error predictions from hippocampal theta power. Calibration replaced ten other-session training rows, so training size stayed equal. Both models scored the same later trials. Public OpenNeuro ds004752 version 1.0.1 supplied 13 patients, 56 sessions, 1,637 scored trials, and 161 errors. The fixed ridge logistic model had penalty 10. Mean patient log loss rose from 0.323783 to 0.329467 nats per trial. The mean gain was -0.005684 nats. Its 95 percent patient bootstrap interval was -0.017688 to 0.005834, under independent patient sampling. Four patients improved. The exact sign-flip probability was 0.409424, under symmetric null patient effects. The test did not establish a calibration benefit or a population increase in loss. No fixed sensitivity test established a benefit. This is a secondary analysis of the same clinical cohort used in earlier work. Training can use later-numbered sessions; calendar order is unverified. The test does not establish prospective deployment performance. The package preserves code, exact compact inputs, source versions, fitted states, probabilities, all planned results, and repeat instructions. Independent raw checks passed at all three sample rates. A second offline run reproduced all 188 analysis files exactly.

SEEG · hippocampus · theta power · working memory · memory errors · session calibration · logistic regression · cross-session prediction