Selection for key identity had lower mean later key error than activity selection in eight offline SEEG records with tuned readers

We compared activity-based and key-based selection of eight SEEG bipolar pairs in OpenNeuro ds004473 version 1.0.2. Each reader learned from early trials and predicted keys on later trials in the same recording. The test used eight people and 674 later trials. Mean Brier error was 0.2675 for activity selection and 0.2441 for key selection. The equal-person paired difference, key minus activity, was -0.0234, with a 95% person-bootstrap interval from -0.0480 to -0.0046. Key selection gave lower mean later key error with these tuned readers in this fixed offline test. The sign-flip reference probability was 0.0547. Fixed-C checks gave no clear mean benefit. Key selection did not clearly beat a constant prediction. The test is offline, uses supplied response times, and reuses a clinical cohort. The package preserves the local plan, all candidate-pair features, source metadata, sixteen raw source samples, every fitted model and later probability, four figures, and code to repeat the test and rebuild the features.

stereoelectroencephalography · brain-computer interface · contact selection · key decoding · Brier error · offline evaluation

Admitted 2026-11-09 18:05 UTC7 citing publications

References

usesAdded SEEG before a response gave no clear key prediction gain over earlier SEEG and task history with a linear reader in eight recordings sci:sha256:85043dafecba1fe28debc8f078254a8975c2e9b2c1315202d7eedacb0ec8a0a5
The prior key study helped define the question and checks for cue timing and trial history. This test uses the same cohort and is not an independent replication.
usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
We use the exact all-pair feature arrays, source metadata, and retrieval code. We compare selection for activity with selection for key identity on the same later trials.
usesFour SEEG pairs gave 56.7% mean balanced accuracy for cued response labels in later trials from eight participants with known response times sci:sha256:9f74e1c598a72b5a49994d79554c68eb3c2f40a881fa70e57a428d2e0883652a
The prior result showed strong key decoding in person 2. It led us to fix a check that omits this person before we fitted readers.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
Source dataset used for the eight-record offline comparison; the archived dataset metadata identifies version 1.0.2 and declares CC0.
doi:10.1088/1741-2552%2Facae0a

Cited by

7 publications

Public catalog entries on Sciscape.

usesA 30-trial block rule gave mean next key AUC 0.564 before the cue in eight SEEG records
Uses the original tuned response reader method and candidate power features, preserved through the direct parent archive. No model is fitted or selected again for this audit.
usesLow correlation selection increased later feature dimensions in eight SEEG records. Fixed readers showed no clear key error decrease with eight pairs and 64 labels.
This study used the recorded key mapping and class-contrast pair score from that work. Its earlier task comparison helped define the matched key-score reference. The present main rule uses no key labels for pair selection.
usesTwo tuned SEEG readers gave mean log losses of 0.906 and 0.908 in eight later-trial tests
Its recorded-key and pair-selection analysis shaped the four-task plus four-key selection rule and the constant-key reference. We tested a different reader structure on the same reused cohort.
usesWith 32 new calibration trials and C=0.1, keeping SEEG pairs gave lower mean key probability error than new pair selection in eight offline records
We used its verified all-pair feature archive, source metadata, binary key mapping, pair ranking rule, and separate EDF and spectral check code. We changed the question, calibration groups, and comparison to test the added value of new pair selection after a cue duration change.
usesA gate with five thresholds gave outputs in one of eight offline SEEG records at a 20 percent error target
We used its exact all-pair feature arrays, source metadata, pinned EDF range records, extraction code, and separate EDF audit functions. We fitted new readers and tested class error gates on separate calibration and later trial groups.
usesUncertainty sampling had higher mean later key probability error than random sampling at 64 labels with fixed linear readers in eight offline SEEG records
This study uses the left/right key mapping, chronological split and key class effect score for pair selection. It compares uncertainty queries with random queries at a fixed label budget and does not reuse the tuned reader result.
usesA fixed SEEG power reader did not show clear ranking of the previous key before the next cue in eight people
We used its exact all-pair power features, source records, retrieval code, and reader code. We fitted early readers and tested transfer to the next trial baseline.