Selected SEEG correction gave higher mean task error forecast loss in 15 additional offline MetaRDK tests
We tested a fixed task error forecast in 15 additional public MetaRDK SEEG records. Early trials selected a task model and zero correction or a correction from eight neural signals. The source marker channel was excluded. The test used 1,502 later trials with 446 errors. Mean person log loss was 0.649877 for task inputs and 0.665339 after selected SEEG correction. The mean gain was -0.015462 natural log units per trial. The descriptive 95% person bootstrap interval was -0.033319 to -0.000755. The exact sign-flip p value was 0.130859. The summaries use different assumptions and do not give the same indication of uncertainty. Three records improved, eight worsened, and four used zero correction. A constant forecast had lower mean loss than both methods. All 537 prior main forecasts matched exactly. A fresh full source rebuild matched one additional record exactly. The package preserves two figures, compact inputs, source versions, code, model states, forecasts, and replay steps. The result applies to this offline method. It does not establish online BCI timing or stimulation benefit.
SEEG · brain-computer interface · task error forecasting · log loss · neural signals · MetaRDK · offline evaluation
References
4- usesEarly validation selected zero SEEG correction in seven of 15 additional MetaRDK records sci:sha256:ebde0b3382e828478c966b9019d5afe3d5022218b1c8a4f0d08d8f40be3859e6
We used all 21 unchanged compact signal records, source behavior and anatomy metadata, source manifests and extraction receipts, and retrieval tools. We used these inputs for task error forecasts. We did not use the source confidence forecasts as error forecasts.
- Eight high gamma SEEG means do not reduce mean forecast loss for later task errors in six MetaRDK records sci:sha256:7beb394715f39fed5df47828831fba10e67541970afc2c613e2757003b311742
[Legacy relation: reproduces] We kept the numerical error forecast method fixed and matched all 537 prior main forecasts exactly. We also matched task, forced, and constant forecasts. We tested the method in 15 additional records, with the marker channel removed, and kept the prior six records separate.
- usesdoi:10.18112/openneuro.ds006253.v1.0.3
The analysis used the 1.0.3 MetaRDK release and unchanged compact records derived from its source data.
- doi:10.1038/s41467-025-61744-8
Background source paper for the reported MetaRDK SEEG acquisition and preprocessing details.
Cited by
3 publicationsPublic catalog entries on Sciscape.
- usesOffline SEEG readers trained on correct trials gave 48.6 percent balanced accuracy for inferred responses before movement on later errors in 15 MetaRDK records
Unchanged compact arrays, behavior and anatomy metadata, source manifest, inherited source receipts, and retrieval tools. Its error forecast outcome is not our response-identity outcome.
- usesSEEG error readers had lower mean AUC across opposite visual cues after movement in 19 offline records
We used its report to check the scope of earlier error forecasts in the same source cohort. It tested selected neural corrections in 15 additional records. We cite it as related work, not as independent evidence for the current transfer result. We did not reuse its fitted models.
- usesEight SEEG means increased offline forecast loss for later changes of mind in all 14 eligible additional MetaRDK records at penalty 10
We used all 21 unchanged compact SEEG records, source behavior and channel metadata, pinned signal versions and hashes, extraction receipts, and retrieval code. A fresh full source rebuild matched sub-19 exactly. We used these inputs for later changes of mind, not for the source publication task error target.