Eight SEEG key readers gave 53.4 percent mean balanced accuracy after condition changes with 32 calibration trials

We tested key reading after a task condition change in eight public SEEG records. Each reader used eight adjacent contact pairs and 32 early calibration trials. The next condition supplied 979 test trials. Mean balanced accuracy was 53.4 percent across people. The descriptive 95 percent person bootstrap interval was 51.4 to 55.5 percent. Mean ROC area was 0.524, with an interval of 0.492 to 0.551. Response power exceeded earlier rest power by 4.5 percentage points. A separate response-versus-rest reader gave 63.0 percent accuracy. Its selected pairs did not give a clear gain for key reading. Using 64 calibration trials did not give a clear gain. Key probabilities had mean Brier error 0.358; a constant probability of 0.5 gives error 0.25. A larger fit penalty reduced error to 0.291 without a clear accuracy gain. Cues matched keys on 960 of 979 test trials. The study cannot isolate motor intent. It uses known later response times and tests neither live commands nor brain stimulation. The package has a local pre-score plan, three figures, all candidate features, trial labels, predictions, model states, source code, pinned retrieval recipes, and hashes. All 48 fits passed independent solver checks. All 2,032 source trial ranges were retrieved again. Retained and recovered arrays matched exactly. This is an exploratory test in a reused clinical cohort.

stereoelectroencephalography · brain-computer interface · key decoding · balanced accuracy · calibration · exploratory study

Admitted 2026-11-09 17:50 UTC9 citing publications

References

usesRemoving the check after the response restored three trials and reduced observed mean later balanced accuracy by 0.043 percentage points in eight fixed SEEG readers sci:sha256:0b35aec3380c9b238ef366d0bb78391c41d801af83381ae5ce7a85c0ab0ad6c7
We used the retained candidate power arrays, recovered arrays and voltage peaks for 18 source rejections, trial metadata, source headers, and 26 saved raw ranges. These inputs permitted a new actual-key test with a voltage rule that uses only pre and rest windows. We fitted all readers again.
usesEight fixed SEEG decoders scored 65.8% mean balanced accuracy for response versus rest windows after task condition changes sci:sha256:27c3a162d5f23c03418feb1337bb50c3dbd6fcf78da3fdc3bdc12da973491084
We used the time-ordered condition test and the response-versus-rest comparison design. We applied new 32-trial calibration and actual response-key targets. No earlier fitted parameters or test predictions entered the new readers.
usesAdding inter-trial SEEG power increased previous-key loss in eight fixed linear later-trial tests sci:sha256:a20461bbd91162cdb6926278933e0f647f3c5da425f809a909270b9dbbf0292c
We used its source-label audit to avoid the field named correct. The new targets use actual keypressed values and each record's saved leftkey and rightkey codes. We do not use the correct field in selection, fitting, or scoring.
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 used the original candidate power arrays, pinned OpenNeuro metadata and range hashes, and source EDF extractor carried through the selection audit. They supplied three power bands and the complete contact-pair candidates. A new full retrieval checked the arrays and all source range hashes.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis uses the forced two-choice SEEG dataset, version 1.0.2.
supportsdoi:10.1088/1741-2552%2Facae0a
The cited source report describes the task and movement-related oscillations.

Cited by

9 publications

Public catalog entries on Sciscape.

usesInput selection by effect agreement changed mean later state Brier error by −0.0008 with 64 trials and eight pairs in eight SEEG recordings
We used its response-key definitions, recovered trials, eligibility rule, and reader code through the immediate source package.
usesTwo 32-trial calibrations within one condition shared 2.5 of eight state pairs and 0.625 key pairs on average in eight SEEG recordings
We use its eight SEEG feature arrays, trial records, recovered features, and source range records. We also adapt its loader and fixed reader. These inputs let us compare maps from separate calibration blocks.
usesKnown simulated contact loss training reduced mean worst-contact Brier error in eight offline SEEG tests with unchanged surviving inputs
We reused its recovered trial inputs and response key eligibility checks through the preserved loader and recovery code.
usesSEEG state maps shared 2.50 of eight pairs between time blocks and 2.24 between matched random groups on average in eight recordings
The inherited input rules supplied recognized key labels and recovered trials. We preserved those rules and checked source trial events.
usesKnown simulated contact loss training reduced mean worst-contact Brier error in eight offline SEEG tests with unchanged surviving inputs
We reused its recovered trial inputs and response key eligibility checks through the preserved loader and recovery code.
usesTwo selected SEEG pairs had higher mean response detection error than eight pairs after condition changes in eight people
We used its candidate SEEG power arrays, recovered trial arrays, event joins, metadata, input loader, and pinned retrieval recipes. These files supplied the inputs and source checks for the new pair budget and probability error tests. We fit all models again and audited one fresh voltage range for each person.
usesA direction rule changed mean key Brier error by -0.00036 for eight SEEG pairs after 64 calibration trials in eight people
We adapt its input loader and fixed logistic reader. We preserve its feature recovery and source retrieval code with the MIT license.
usesConfidence from other people lowered probability error after condition changes in eight response-aligned SEEG key readers with 32 training trials
Used the complete original and recovered SEEG power arrays, trial eligibility rules, 32-trial key reader implementation, metadata, and pinned raw-range retrieval code. Rebuilt the arrays and refitted the fixed reader for a new donor confidence test.
usesCurrent rest subtraction gave no clear log loss gain in eight offline SEEG key readers after task changes with 64 calibration trials
This study uses the corrected eight-patient input cohort, source feature arrays, recognized-key joins, and retrieval code. It tests rest subtraction and forward penalty selection with key probability loss.