Correction with unlabeled rate estimates gave higher Brier error than known rate correction in eight SEEG tests at a simulated 10 percent task rate

We tested an unlabeled rate estimate for probability correction in eight public SEEG recordings from OpenNeuro ds004473 version 1.0.2. An earlier sci result used a supplied rate. We test the missing estimate. Four trial blocks separated model fitting, calibration, unlabeled rate estimation, and final testing. At an assumed 10 percent task rate, known rate correction gave mean Brier error 0.0762. A soft class mean estimator used 200 unlabeled windows in each of 200 simulations per recording. Correction with these estimates gave mean Brier error 0.1006. Known rate correction had less error in all eight recordings and beat the constant probability 0.10 in all eight. One weak score separation triggered the planned 0.50 fallback. Without that recording, the descriptive mean Brier penalty was 0.0083. The primary later-pool minus matched-control absolute rate error was 0.0333; the exact two-sided p value was 0.125. The exact test was inconclusive. Score mean changes and weak separation limited the tested estimator. The rate is a simulation assumption, not a measured command rate. The package preserves features, models, draws, code, three figures, and source checks for 64 public EDF trial ranges. This is an offline task state result. It does not test live BCI control or stimulation.

brain-computer interface · SEEG · label shift · probability calibration · Brier error · class rate estimation · offline benchmark

Admitted 2026-11-09 19:06 UTC8 citing publications

References

usesA scale from other patients gave mean Brier error 0.187 in eight offline SEEG tests sci:sha256:bc9061e8120e271886ed956e358f68c3c9ea381de284673bc87b058f6d467a7c
We use its positive score scale procedure and fixed bounds of 0.05 to 20. Our calibration block and fitted models are new.
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 sci:sha256:d170a3167e027151af000cd062b631f63de7fd71e97331686d0550ec09f7c42a
Its correction with a supplied rate and stated need for a rate estimate define the next question. We test a soft class mean rate estimate on four new trial blocks. We do not reuse its test predictions.
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 its public SEEG features, trial records, source metadata, and range retrieval code. We fit new models on four separate trial blocks.
usesRest calibration gave a 4.2 percent mean false positive rate in eight later-trial SEEG tests sci:sha256:7a28b8fc519f8f608fe5546ba40889db9bb37b5c93f68ed3821b4ba5bf1c5e75
Its later rest calibration result motivated a test of probability error when task events are rare. We test an unlabeled rate estimate with separate adaptation and final test blocks.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analyzed public SEEG dataset is OpenNeuro ds004473 version 1.0.2.
doi:10.1088/1741-2552%2Facae0a
doi:10.1162/089976602753284446

Cited by

8 publications

Public catalog entries on Sciscape.

usesConstructed rest score changes gave equal score distributions at 1 and 10 percent task rates for eight offline SEEG readers
We use the eight fixed readers, raw scores, chronological splits, feature arrays, and source records. We reconstruct all raw scores and test CDF rate bounds without new model fitting.
usesUnlabeled class rate updates from 12 trials showed no clear Brier error decrease in 15 SEEG session tests at C = 0.1 under assumed timing
This study supplied the related question about rate estimates without labels. The present test used actual numbered sessions in another dataset and a maximum likelihood objective. It did not repeat the earlier simulated rate experiment.
usesScale fitting at a supplied 1 percent task rate reduced later log loss in only two of eight SEEG tests compared with fitting at 50 percent
This work provides the SEEG rate-correction context. The present test compares two scale-fitting losses with the same supplied rate. It does not reuse fitted probabilities.
usesScale fitting at a supplied 1 percent task rate reduced later log loss in only two of eight SEEG tests compared with fitting at 50 percent
This work provides the SEEG rate-correction context. The present test compares two scale-fitting losses with the same supplied rate. It does not reuse fitted probabilities.
usesEight SEEG tests did not establish 50 percent precision under binomial bounds that assume independent trials, stable class distributions, and a 1 percent task rate
Its probability analysis at an assumed 10 percent task rate provided context for a reader with a low task rate. It motivated the new test of separate threshold calibration and finite-sample precision bounds at an assumed 1 percent task rate.
usesEight SEEG tests did not establish 50 percent precision under binomial bounds that assume independent trials, stable class distributions, and a 1 percent task rate
Its probability analysis at an assumed 10 percent task rate provided context for a reader with a low task rate. It motivated the new test of separate threshold calibration and finite-sample precision bounds at an assumed 1 percent task rate.
usesScores reduced mean Brier error with eight adaptation labels in eight offline SEEG tests at a simulated 10% task rate
The direct parent supplied feature files, source records, model code, and fixed trial blocks.
usesScores reduced mean Brier error with eight adaptation labels in eight offline SEEG tests at a simulated 10% task rate
The direct parent supplied feature files, source records, model code, and fixed trial blocks.