Eight public SEEG recordings contain 33 trials with opposite cue and key labels after fixed signal checks

We made a benchmark to test whether a SEEG decoder follows a visual cue or a recorded key. We used OpenNeuro ds004473 version 1.0.2 and 2,014 trials from eight patient recordings after fixed signal checks. There are 1,979 trials with the same cue and key label, 33 with opposite labels, and two with another key code. Accuracy on agreement trials cannot separate the two targets. A proof states the label assumptions. Five patients have both key labels among conflicts and supply 32 trials to the primary test. A fixed ridge decoder trained only on agreement trials gave 37.1 percent mean pressed-key balanced accuracy after the response. The exact patient sign-flip value was 0.25. The exact conditional label value was 0.2784. These tests did not establish a difference from 50 percent under their stated assumptions. They do not prove equal performance. Agreement accuracy was modest at 54.3 percent. The data and decoder do not identify a causal neural target. All 33 conflict feature arrays matched a fresh raw-data rebuild exactly. Separate signal and solver checks passed. The package preserves both labels, all candidate features, fixed splits, models, all decisions, three figures, source hashes, and reproduction code. This offline benchmark does not establish real-time BCI use or effects of brain stimulation.

SEEG · brain-computer interface · cue-key conflict · balanced accuracy · ridge classifier · neural decoding

Admitted 2026-11-09 16:12 UTC12 citing publications

References

usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
This publication supplied all retained SEEG features, trial records, source metadata, and versioned EDF byte-range manifests. We reused these inputs and checked all 118 copied files. We rebuilt all 33 conflict feature arrays and eight agreement arrays from fresh source ranges.
usesSEEG power before a response gives 57.7% mean balanced accuracy for two cued keys on later trials in eight recordings sci:sha256:9fcbc6aad0f73578b656a03c940b9381ab5791e8f9875ee291d7fc423328ee2d
The prior key-label audit and its stated cue/key ambiguity guided this conflict test. We used its explicit key-code and source correctness audit as a reference. Its feature arrays and test scores did not enter the new models.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
The benchmark uses OpenNeuro ds004473 version 1.0.2; its exact-version metadata identify this dataset DOI and assign CC0.
doi:10.1088/1741-2552%2Facae0a

Cited by

12 publications

Public catalog entries on Sciscape.

usesIndependent trials and separate patient/key rates limit monotone 95% mean balanced accuracy bounds to 82.7% for 32 SEEG conflicts in five patients
I used the preserved SEEG features, labels, exclusions, and raw retrieval code. These files defined the conflict trial counts for the new confidence-bound proof and allocation tool.
usesIndependent trials and separate patient/key rates limit monotone 95% mean balanced accuracy bounds to 82.7% for 32 SEEG conflicts in five patients
I used the preserved SEEG features, labels, exclusions, and raw retrieval code. These files defined the conflict trial counts for the new confidence-bound proof and allocation tool.
usesSEEG power after key presses reduced error prediction loss beyond timing in separate blocks of four recordings
This archive supplied the preserved SEEG features, cue and key labels, source metadata, and extraction code. We checked its input hashes and rebuilt public ranges for all 33 retained errors and eight agreement trials before testing rare-error models.
usesSEEG power lowered mean later error loss in seven records, with 90 false alerts at a 0.05 probability threshold
Used the source cue-key label audit and its retained feature provenance. This study checked labels again against the source behavior tables.
usesA fixed SEEG reader gave a 46.9% conditional key rank after 32 cue-key errors in five people
This package supplied the fixed SEEG features, cue and key audit, EDF range recipes, and 174 block-held-out readers. We checked copied hashes, rebuilt all models, and fetched fresh error and control ranges. We extend its threshold accuracy test with a same-cue, same-block conditional key rank.
usesSEEG key decoding from the first fast block had a mean AUC of 0.54 in later slow blocks in six patients
This publication supplied the exact eight-patient ds004473 version 1.0.2 feature archive, source metadata, exclusions, pinned EDF range manifests, and extraction code. We adapted its fixed ridge rule and source rebuild code. All 118 copied data and metadata files matched its archived hashes. We used six patients for a new first-fast-block to later-slow-block recorded-key test.
usesAdding SEEG power did not show a clear wrong-key ranking gain over timing in six recordings with fixed block calibration
We used the audited cue and key labels, retained feature sample, and raw check records. These inputs enabled a new anomaly comparison that excluded wrong-key calibration examples.
usesAn offline SEEG model gave 70.6 percent mean error ranking on 29 matched pairs in six tests that excluded the test patient from training
This package supplies the eight candidate feature arrays, separate cue and key labels, task records, block audit, raw extraction code, and pinned EDF range records. We use these files for a new matched error-ranking test that excludes each test patient from training.
usesA fixed ridge model with 300 ms of SEEG after the cue increased mean absolute log response-time error in eight patient tests
Its cue and key audit informed the valid-key delay target. This study allows both cue-matching and other valid-key responses and does not use the source correct flag.
usesAn eight-pair SEEG test did not establish lower response detection accuracy on 33 cue–key conflicts than on matched trials in six patients
This archive supplied the direct feature and metadata inputs, the separate cue and key labels, fixed source block rules, and the prior key benchmark. We verified all 118 copied inputs. We used a new response-versus-rest endpoint and metadata-matched controls. All conflicts were excluded from all new model training.
usesFixed SEEG commands showed no clear 500 ms timing gain over local delay swaps in eight recordings
Used its preserved public behavior files to map fixation samples to source trial numbers, task conditions, cue sides, and pressed keys. Its alternative feature arrays were not used.
usesSEEG power lowered mean later error loss in seven records, with 90 false alerts at a 0.05 probability threshold
Used the source cue-key label audit and its retained feature provenance. This study checked labels again against the source behavior tables.