Encoding power reduced later recall prediction error after history control in seven Utah patients under fixed ridge models

We tested whether encoding power reduced recall prediction error after control for three past released recall counts. We used public recordings from OpenNeuro ds006136 version 1.0.1. The main group had seven Utah patients with reported depth implants. Individual channel construction remains uncertain. The release identifies the recordings as SEEG. Fixed ridge models trained on earlier trials only. The test scored 846 later trials. Added beta and high gamma power reduced mean patient error from 0.543240 to 0.528530 squared count units. The gain was 0.014710, or 2.71%. All seven patients improved. The 95% patient bootstrap interval was 0.005810 to 0.027540. The exact sign assignment test gave p = 0.015625. It assumes independent patient gains and a symmetric zero null. History features themselves increased mean error. This result supports one fixed model comparison, not a memory mechanism or clinical use. All five computed checks had positive mean gains. A wider training gap lacked enough rows for two channels and was not computed. The package preserves code, source recipes and hashes, input features, all predictions, main fits, resampling arrays, figures, and independent checks. A fresh feature rebuild and a repeat numeric run were byte-identical.

encoding power · recall prediction · working memory · ridge regression · intracranial EEG · patient bootstrap