Block tests of fixed models did not show a clear recall prediction gain from delay power in SEEG data from 13 patients
We tested whether delay beta and high gamma power improves recall count prediction beyond fixation power and trial position. We used OpenNeuro ds006136, version 1.0.1. The data include 13 patients, 14 sessions, 143 channel records, and 2,134 original trials. We matched channels by original trial ID. We kept all released channels and all recall counts. The plan fixed ridge models, signal windows, a two-ID gap, and five test blocks before power extraction. Each patient had equal weight. Mean square error fell from 0.417746 to 0.408972 recalled objects squared. The mean gain was 0.008774, with a 95% patient bootstrap interval of -0.002455 to 0.020765. The exact sign-flip probability was 0.192383 under a symmetric patient-effect null. The primary result did not show a clear patient mean gain. The fixed secondary forward test gave a gain of 0.014675 on 410 later trials. These results do not establish an absence of useful memory signals. All source hashes, padding checks, and cross-channel score checks passed. Equal packed row positions had different original IDs in 87.60% of channel-pair comparisons. Independent numerical checks passed. A full rebuild gave byte-identical data and model files. The package preserves code, source metadata, fixed source versions, power features, folds, fits, predictions, patient results, figures, and reproduction instructions. Raw signals have an exact retrieval recipe. The scope is accepted trials within these sessions. It does not cover new patients or clinical use.
working memory · stereoelectroencephalography · recall prediction · band power · ridge regression · cross-validation · reproducibility
References
2- usesdoi:10.18112/openneuro.ds006136.v1.0.1
Source signals, original trial IDs, behavioral scores, and metadata from OpenNeuro ds006136 version 1.0.1.
- doi:10.1016/j.neuroimage.2026.121718
Cited by
5 publicationsPublic catalog entries on Sciscape.
- usesEncoding band power reduced mean blocked recall prediction error across 13 patients from OpenNeuro ds006136
The earlier count benchmark supplied the patient endpoint and fixed ridge penalty design. This study tests encoding power with an available channel control and different trial blocks.
- usesAdding encoding high gamma increased mean relative late delay power prediction error in fixed models across trial blocks from 13 SEEG patients
This work supplied the pinned signal manifest, retrieval code, source metadata, and extraction starting point. We retrieved the signals again and computed new encoding and delay features for a measured-power target.
- usesDelay high gamma power concentration gave no clear recall prediction gain in fixed block models from 13 SEEG patients
We read this earlier trial block recall result before our analysis. It supplied context for comparison with mean delay power. Our added feature measures uneven delay power across time segments and uses new signal features from the same patients.
- usesA fixed channel fraction model gave no clear recall prediction gain across blocks in 13 SEEG patients
We used its channel feature table, source metadata, pinned signal manifest, and retrieval and extraction code. We rebuilt the availability table and fitted new session models. The cohort is shared.
- usesEncoding band power reduced mean blocked recall prediction error across 13 patients from OpenNeuro ds006136
The earlier count benchmark supplied the patient endpoint and fixed ridge penalty design. This study tests encoding power with an available channel control and different trial blocks.