Calibration trial reuse for reader fitting reduced mean trial coverage by 5.5 percentage points on later trials in eight SEEG recordings at a 90% target
We tested label sets for rest and pre-response states in public SEEG data from eight people. At a 90% trial coverage target, reuse of calibration trials for reader fitting lowered mean coverage from 91.1% to 85.6% on later trials. Coverage fell in all eight people. The mean loss was 5.5 percentage points, with a 95% bootstrap interval of 2.5 to 9.7 points. Raw reader accuracy rose from 70.5% to 71.3%. The loss was 6.0 points with random trial allocation. The package includes a four-method benchmark, figures, saved models, trial assignments, source checks, and code. This test uses one task in eight clinical recordings. It does not test live control or stimulation.
stereoelectroencephalography · brain-computer interface · split conformal prediction · calibration data reuse · label sets · trial coverage · logistic regression · neural signal decoding
References
5- SEEG 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 sci:sha256:8caefc45395d2f1a318b4c028f4a3521b30eabfe8cebe523db5f70859263a998
[Legacy relation: reproduces] Used the same separate-data splits and joint trial score. Exactly reproduced 32 joint-set result rows at shared targets. Added controlled reuse of calibration trials for reader fitting and contact selection.
- usesFour-bit error feedback reduced mean state log loss at 999 Hz and increased it at 250 Hz in eight SEEG recordings with fixed readers and limits of six times training RMS sci:sha256:91e54b754ed955d5fa0916d3e8b3aa5659aeb3b71f70125655aa74a6b185a587
Used the MIT-licensed numerical helper for logistic fitting, scoring, patient bootstrap intervals, and paired sign-flip statistics. This publication does not test signal quantization.
- usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
Used the exact candidate-pair feature arrays, eligibility rules, source sidecars, pinned retrieval recipes, and feature extraction code. All 118 inherited files matched the archived hashes.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis uses the eight SEEG recordings in OpenNeuro ds004473 version 1.0.2.
- doi:10.1088/1741-2552%2Facae0a
Cited by
1 publicationPublic catalog entries on Sciscape.
- usesA 20-trial delay in correct feedback changed mean absolute coverage error by 0.23 percentage points in eight SEEG replays at a 90% target and a 0.05 step
Used the separate C=1 reader, exact trial splits, double precision probabilities, saved model parameters, and source provenance. The new analysis adds delayed correct-label feedback. The fixed-set counts matched the parent.