Trial ensembles reduce SEEG probability error with eight fixed pairs after 32 calibration trials in eight recordings

We tested whether trial ensembles reduce SEEG probability error under an eight-pair input limit. All eight patients from OpenNeuro ds004473 version 1.0.2 were used. Each reader learned from the first 32 retained trials and was tested on the final third. The test had 674 trials. A mean of 256 readers with the same eight pairs reduced equal-patient mean Brier error from 0.2817 to 0.2550. All eight reductions were positive. The paired reduction was 0.0267, with a descriptive 95% person interval of 0.0172 to 0.0359. Balanced accuracy changed from 62.61% to 63.36%. A fixed half-confidence rule gave lower mean error, 0.2299, with no new fits. Selecting pairs again in each member gave error 0.2102, but required 59–98 distinct pairs and 80–121 contacts. A rest time control also showed lower probability error near chance accuracy. The study supports input-count and confidence controls for offline BCI reader tests. It does not establish continuous control or stimulation benefit. Code, features, source records, model states, predictions, checks, and figures are preserved.

SEEG · brain-computer interfaces · trial ensembles · Brier error · probability estimates · input budget · offline decoding

Admitted 2026-11-09 18:48 UTC4 citing publications

References

All available SEEG pairs had an uncertain mean task-state accuracy gain over eight pairs after 32 calibration trials with L2 C=1 in eight records sci:sha256:2dfb88b8b49b59c7ed25acb61dd560d18def2a21039909a1de4bbe2388d33b0c
[Legacy relation: reproduces] The single eight-pair and all-pair readers repeat the C=1 first-32-trial comparison on the same final-third test trials. Their equal-patient accuracy values match the prior 62.6% and 64.2% results. They serve as checks and baselines for the new probability and input-budget comparisons.
usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
The study uses the complete eight-patient feature archive, trial and exclusion records, pinned source metadata, raw audit ranges, and retrieval and extraction code. These files supply all inputs and source checks for the new ensemble tests.
usesTrial resampling reduced offline error at 50 percent acceptance in eight SEEG recordings with 32 calibration trials sci:sha256:761ec71bb83cda1613335fce20a94d09f5c01d0eee0ceaf99b7c786619757ab9
Newton fitting, independent BFGS fitting, scalar EDF reads, and direct Fourier checks were adapted from this publication. Its trial resampling method informed the paired trial draws. Fitted states and predictions were not reused. The new test averages member probabilities and measures the combined input count.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
The study analyzes the eight recordings from this exact OpenNeuro dataset release.
doi:10.1088/1741-2552%2Facae0a
doi:10.1007/BF00058655

Cited by

4 publications

Public catalog entries on Sciscape.

usesContact selection before cross validation raised estimated balanced accuracy by 15.6 percentage points in eight SEEG recordings with 32 calibration trials
Uses the fixed half confidence rule and the 32-trial fixed reader as pre-specified comparisons.
usesAn eight-pair reader trained from SEEG ensembles reduces mean Brier error compared with half confidence in eight recordings
We use its frozen early-trial ensembles and copied MIT code. We transfer their probabilities into new readers and test final input costs.
A fixed resampling selector raised mean later Brier error in eight SEEG recordings with 32 calibration trials and eight input pairs
[Legacy relation: reproduces] Our independent run reproduces the standard eight-pair reader Brier error of 0.2817 and the half-confidence control error of 0.2299 on the same 674 later trials. We do not recompute its ensemble.
usesA fixed resampling selector raised mean later Brier error in eight SEEG recordings with 32 calibration trials and eight input pairs
Its ensemble with repeated pair selection required 59–98 distinct pairs. That input-budget result motivates our new test of one frequency-selected eight-pair panel.
usesWorst-loss training gave mixed later error changes compared with mean-loss training under known single-shaft loss in eight SEEG recordings
The study uses the fixed half-confidence rule p_new = 0.5 + 0.5*(p - 0.5) as a control. It uses the same 32-trial training check and does not fit this rule to later labels.