Early confidence scaling decreased mean Brier error from 0.193 to 0.185 in eight offline SEEG patient tests

A brain-computer interface needs useful confidence values as well as correct class decisions. We tested one confidence scale on public OpenNeuro ds004473 version 1.0.2. Each of eight patients had an eight-pair SEEG decoder. Early training, separate early calibration, and later testing used different trials. The test used 674 trials and balanced response and rest windows. Mean Brier error decreased from 0.19326 to 0.18473. Five patients improved and three became worse. The paired mean change was -0.00853. The 95 percent patient bootstrap interval was -0.01921 to 0.00094. The exact sign-flip p value was 0.164. Mean balanced accuracy stayed at 70.5 percent. Scaling on reused training data increased later Brier error in all eight patients. A rest-position control also had lower probability error after scaling, while its accuracy stayed near chance. The package preserves features, trial predictions, four figures, source records, and code. Independent solver checks, an exact repeat, and fresh public raw ranges passed. Task conditions changed across time and limit interpretation. This small offline test does not establish benefit across patients or measure continuous intent, live command risk, or brain stimulation.

stereo-EEG · brain-computer interface · probability calibration · Brier score · temperature scaling · electrophysiology

Admitted 2026-11-09 17:50 UTC11 citing publications

References

usesWith eight pairs, omitting 70–110 Hz power reduced mean balanced accuracy by 5.6 points in eight offline SEEG tests sci:sha256:e3688bef3ed8761381d86408e300703e855ca508766aef3507db1ef78561ef10
We use its complete eight-patient SEEG power features, trial exclusions, source records, raw retrieval code, and decoder design. We add a separate early confidence fit and a later probability test.
usesA rest threshold set for 5 percent errors gave 3.9 percent false detections and 26.5 percent sensitivity in eight later-trial SEEG tests sci:sha256:2100f971aed9c12a7d4be6af5c0b8c155e730ec5176d7b8c6dd80fcfd083230d
We use its preserved complete raw-range reconstruction record as inherited evidence for the feature inputs. This run adds partial independent raw checks. It does not repeat the full reconstruction.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
The publication analyzes the eight SEEG recordings from this exact OpenNeuro version.
doi:10.1088/1741-2552%2Facae0a

Cited by

11 publications

Public catalog entries on Sciscape.

usesTemperature scaling and prior correction lower Brier error below a constant forecast under a 5% response weight in later task windows from eight SEEG recordings
This study uses the separate model, temperature, and later test split. It fits new models and tests weighted response fractions.
usesTrial resampling reduced offline error at 50 percent acceptance in eight SEEG recordings with 32 calibration trials
Its probability scaling result motivated a test of error rejection at matched acceptance. Positive temperature scaling preserves ranks. The new trial-refit score can change ranks. We do not reuse its fitted models or code.
usesSEEG sets at a 90% trial coverage target gave a mean output yield of 34.2% and a mean accepted error of 17.3% in eight later-trial tests
We used its separate fit, calibration, and later test split to measure label sets and accepted output error.
An early scale reduced log loss for fixed Gaussian readers in 10 SEEG patients and 6 fMRI adults
Related early SEEG confidence test with mixed Brier results, read through its catalog abstract.
usesOffline SEEG label sets had 20.2 percent mean single-label error at a nominal 90 percent coverage target in eight recordings
This result supplied the confidence-score context and temporal split used in this test. The present study tests separate class label sets and error among single labels. Both studies use the same public recordings, so they are not independent replications.
usesA scale from other patients gave mean Brier error 0.187 in eight offline SEEG tests
We used its compact inputs, source records, fixed decoder choices, and chronological split. We read its target-only calibration results before this exploratory test. We tested a new scale that uses only other patients.
usesTemperature scaling and prior correction lower Brier error below a constant forecast under a 5% response weight in later task windows from eight SEEG recordings
This study uses the separate model, temperature, and later test split. It fits new models and tests weighted response fractions.
usesAn added voltage screen after a key press reduced input coverage in eight SEEG records at the training rest 95th percentile
We adapted its raw range retrieval code and Newton logistic solver. We used its source metadata and range hashes. We restored the 18 trials removed by its voltage rule.
usesConfidence from other people lowered probability error after condition changes in eight response-aligned SEEG key readers with 32 training trials
Used the separate confidence fitting set and constant-probability control rationale from this prior response-versus-rest confidence study to specify a new key identity and other-person confidence question. No fitted outputs from that study enter the present scores.
usesOffline SEEG label sets had 20.2 percent mean single-label error at a nominal 90 percent coverage target in eight recordings
This result supplied the confidence-score context and temporal split used in this test. The present study tests separate class label sets and error among single labels. Both studies use the same public recordings, so they are not independent replications.
usesMaps from 16 trials in a new task condition gave 0.028 bits more mean loss than matched earlier maps in eight fixed SEEG decoders
Its separate early confidence test in the same public cohort informed the new probability question. We compare matched 16-trial maps before and after a recorded condition change, with frozen decoder states. We do not reuse its fitted maps or scores.