Fixed early confidence thresholds reduced mean task state error to 24.5% at 49.6% retention after condition changes in eight offline SEEG tests
A brain reader can reject outputs with low confidence. We tested a fixed rejection rule in eight public SEEG recordings from OpenNeuro ds004473, version 1.0.2. We reused the same trials and fixed readers from an earlier publication. Each threshold was the confidence median from separate early data. The thresholds used 418 held-out trials in the first recorded condition. The next condition supplied 975 test trials. Mean task state error fell from 34.2% on all windows to 24.5% on accepted windows. Mean retention was 49.6%. Each person had lower accepted error. The mean reduction was 9.7 percentage points. Its descriptive 95% person bootstrap interval was 6.6 to 12.3 points. Equal weights for the two accepted states gave a 9.2-point reduction. The return check in three people did not reduce mean error. These saved event windows do not measure live command errors or stimulation control. Condition and time changed together. The inherited trial screen also used samples after the response. The package preserves source features, fixed models, predictions, figures, code, and independent checks.
selective classification · confidence threshold · rejection rule · stereo-EEG · brain-computer interface · condition shift · offline evaluation
References
5- usesRemoving the test correction did not give a 10 percent error gate in eight offline SEEG recordings under a binomial trial model sci:sha256:09e14eac558a49e4a7ce5ae129799ee696dcadb2023ee12404803ea3c8a47851
This strict error gate study defined a relevant limit. We tested mean error reduction and did not claim a certified error bound.
- usesEight fixed SEEG decoders scored 65.8% mean balanced accuracy for response versus rest windows after task condition changes sci:sha256:27c3a162d5f23c03418feb1337bb50c3dbd6fcf78da3fdc3bdc12da973491084
We used this archive for all eight feature arrays, source metadata, primary fixed readers, and split records. It enabled a separate early threshold calibration and a later condition test.
- usesTrial resampling reduced offline error at 50 percent acceptance in eight SEEG recordings with 32 calibration trials sci:sha256:761ec71bb83cda1613335fce20a94d09f5c01d0eee0ceaf99b7c786619757ab9
This batch rejection study motivated the confidence question. We tested thresholds fixed with separate early data. We did not reuse its resampling score or compare its different reader as a matched baseline.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
Source SEEG data and metadata for the eight analyzed recordings; version 1.0.2 is identified in the publication and its archived source metadata.
- doi:10.1088/1741-2552%2Facae0a
Cited by
9 publicationsPublic catalog entries on Sciscape.
- usesA score interval gate kept accepted labels fixed under simulated constant voltage loss of up to 6 dB on one shaft in eight SEEG readers
We reused all eight feature arrays, frozen primary readers, split roles, early confidence thresholds, source metadata, and intact predictions. These records define the paired gate test.
- usesFixed signal distance gave no clear Brier score gain over confidence in eight offline SEEG condition changes
We used its fixed readers, SEEG power features, trial roles, raw audit samples, and source code. We preserved the source files and checked their hashes. Its early confidence gate supplied the reference result.
- usesA 25% random trial audit rate and a delay of five trials increased mean error exposure by 23.9 points against complete immediate feedback in eight offline SEEG tests
We use the complete unchanged feature package, fixed primary reader values, calibration split, median confidence gates, public source metadata, raw range recipes, and independent reader checks. We reconstruct 975 test trial paths and add random audits with fixed feedback delays.
- usesA ±6 dB shaft gain bound retained 9.0% of windows after simulated loss in eight SEEG readers with fixed equal contact gains
The early median confidence thresholds and split method come from this study. The package includes verified copies of its source code.
- usesThree tapers increased mean SEEG gate score at error cost 4 in eight reused recordings, but the change interval included zero
Its early confidence-gate test motivated the output decision question. We compared spectral estimates under a fixed error cost, without using its models or thresholds.
- usesA confidence margin blocked new output errors and withheld most correct outputs in eight SEEG records under an ideal ±3 dB shaft gain model
This publication supplies the early confidence gates through the direct source package.
- usesA 10% conditional error test stopped five of eight fixed SEEG readers in an offline replay with complete trial feedback
We use the complete archived feature inputs, fixed primary reader weights, calibration split, median confidence gates, public source receipts, and independent validation code. We add trial feedback monitoring and stopping replay without a new reader fit.
- usesAgreed task labels from separate SEEG shaft groups had 31.0% mean error after condition changes in eight recordings
This fixed confidence study supplied context for testing output acceptance. We made a new agreement rule and confidence benchmark without reusing fitted states.
- usesA simulated 6 dB shaft voltage loss raised mean accepted label error by 10.5 points in eight SEEG readers with fixed early confidence gates
We reused its eight feature sets, frozen primary readers, split, early thresholds, source metadata, and intact predictions. Its archived values define the paired gain test.