A fixed SEEG recall reader with scaling from 60 early words had higher mean loss than task inputs in eight new session pairs at C = 0.01

We tested a fixed recall reader in eight additional people from OpenNeuro ds004809 version 2.2.0. Each person supplied 240 source words, 60 later calibration words, and 180 later test words. The neural inputs were 96 log power features from 16 bipolar SEEG pairs. We set the later neural mean and standard deviation from the 60 early words. At C = 0.01, mean log loss was 0.5913 after early label calibration. This loss was 0.0190 below source scaling and 0.0114 above task inputs. The 95% person bootstrap interval for the normalization gain was -0.0003 to 0.0469. The paired sign reference was 0.1719. Scaling improved loss in five people and beat task inputs in three people. The eight previously studied people formed a separate secondary group. A numerical check found 24 constant source features with false small standard deviations in one new session. Exact constant detection removed extreme forecasts. All 144 final model files stayed unchanged when test outcomes and test neural inputs were altered. The package has features, raw witnesses, forecasts, code, figures, source versions, and repeat instructions. This offline study does not test stimulation effects.

stereoelectroencephalography · memory recall · session transfer · normalization · logistic regression · brain-computer interface

Admitted 2026-11-09 16:39 UTC5 citing publications

References

usesA fixed SEEG reader had mean recall AUC of 0.547 for random words and 0.544 for complete lists in eight sessions sci:sha256:853393b59d05a011f54cf6ca9a8e345e41914f3d36a0cde98355d02f5daa1c3d
We used the fixed SEEG feature design from this study through the directly cited later package: 16 bipolar pairs, six log power bands, and the one-second window that starts 0.3 s after word onset.
usesAdded SEEG power increased recall forecast loss in all eight later sessions with a fixed logistic reader sci:sha256:5d98ad3a4b35585f15bfec73be164b2de5da77c70912a6e3c5282178fbe7d256
We used its feature files for eight people and its code for EDF retrieval, task coding, offsets, and scores. Its prior transfer results informed C = 0.01 before the eight additional people were selected. The original eight people remain a separate secondary group.
usesdoi:10.18112/openneuro.ds004809.v2.2.0
The analysis used OpenNeuro ds004809 version 2.2.0; its recorded license is CC0.
doi:10.1038/s41467-017-02753-0

Cited by

5 publications

Public catalog entries on Sciscape.

usesOne fitted neural weight per session reduced mean log loss against a fixed weight in 16 reused SEEG session pairs at C=0.01
This work supplied the source task design and normalization rule. We kept these rules fixed to test the new neural weight with the same model inputs.
The mean non-recall rate was 70.1% for a sequential SEEG and task rule with three selections per list in 16 later sessions
[Legacy relation: reproduces] We refitted all 64 archived readers as verification after we calculated the new policy outcomes. The parameters matched within floating-point precision. The policy test used the fixed archived readers. This corrects the order stated in the uses citation: the verification refits did not precede the new policy test.
usesThe mean non-recall rate was 70.1% for a sequential SEEG and task rule with three selections per list in 16 later sessions
We reused all 16 paired-session feature sets, fixed readers, source metadata, and validation records. The prior reader result informed C = 0.01 and early neural scaling. We refitted the readers before testing new sequential selection rules.
usesAn added SEEG model gave an uncertain loss gain in 16 later sessions at C=0.01
This package supplied all 16 selected SEEG session pairs, power features, word labels, source metadata, pinned raw ranges, raw witnesses, and fixed reference model parameters. We used its task inputs and calibration objective.
The mean non-recall rate was 70.1% for a sequential SEEG and task rule with three selections per list in 16 later sessions
[Legacy relation: reproduces] We refitted all 64 archived readers as verification after calculating the new policy outcomes. Parameters matched within floating-point precision. Later test inputs and labels did not change the fits.
usesThe mean non-recall rate was 70.1% for a sequential SEEG and task rule with three selections per list in 16 later sessions
We reused all 16 paired-session feature sets, fixed readers, source metadata, and validation records. The prior reader result informed C = 0.01 and early neural scaling. The new policy test used these fixed readers.
One calibrated SEEG score gave an uncertain recall forecast gain over task calibration in 16 later sessions
[Legacy relation: reproduces] The task-offset and joint-reader loss and Brier scores reproduce 32 parent score rows within 7e-13. The new AUC uses an explicit 1e-12 tie tolerance, so this relation does not claim exact AUC reproduction.
usesOne calibrated SEEG score gave an uncertain recall forecast gain over task calibration in 16 later sessions
This study uses all 64 feature and trial files, source records, selection rules, retrieval code, task coding, and 16 joint-reader model files from this package. It adds a source residual score and matched early task calibration.