Joint score bounds gave 96.3% mean trial coverage and 34.5% mean output yield in eight later SEEG tests with known simulated loss of one contact
We tested score bounds for a reading brain interface. We used public SEEG data from eight patients in OpenNeuro ds004473 version 1.0.2. Each fixed decoder used eight bipolar pairs. Earlier trials supplied the fit and calibration data. Tests used 674 later trials and 105 known single contact loss cases. The target was the intact decoder score. At a 95% target, joint calibration gave 96.3% mean coverage across patients. Separate calibration gave 71.2%. Mean output yield fell from 46.9% to 34.5%. Four patients had joint coverage below 95%. Joint gates emitted 6,277 of 17,682 window and fault cases. Four emitted decisions differed from the intact decoder. However, 1,108 emitted decisions had the wrong task window label. Thus, decision agreement did not establish task accuracy. At the 99% target, finite-sample ranks gave infinite bounds and zero output yield. The proof requires exchangeable calibration and new trial vectors. Task conditions changed across time, and these data do not establish that assumption. The later trials were used by earlier work, so the new tests are exploratory. The package includes an observed-input score tool, a conditional proof, all features and scores, three figures, source records, raw audit checks, and an exact replay.
stereoelectroencephalography · brain-computer interface · conformal prediction · contact loss · score bounds · selective prediction
References
6- usesTraining covariance increased mean balanced accuracy by 0.36 percentage points after known simulated loss of one contact in eight offline SEEG decoders sci:sha256:def7ff6139316b5d2345596c08c24a0bc6d86abe4fa6a51d420c77ccb5674e3d
This work supplied the fixed feature and source files, raw extraction code, and covariance readout code used in this study.
- usesJoint calibration reduced rest windows with any call across intact and known simulated single contact loss cases in eight offline SEEG tests sci:sha256:908f59ed7554b3699ceb71b6c9ccbf84c41f99043720a0086e13707449633e72
This work used a maximum over contact loss cases for rest-call calibration. We adapted that design to intact score error.
- 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 sci:sha256:8caefc45395d2f1a318b4c028f4a3521b30eabfe8cebe523db5f70859263a998
This work used whole-trial label sets and measured task errors among accepted outputs. It guided the trial grouping and error reporting.
- usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
This work defined the source windows and spectral features used through the recovery parent.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The study analyzes SEEG recordings from OpenNeuro ds004473 version 1.0.2.
- doi:10.1088/1741-2552%2Facae0a
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesSeparate pair bounds covered 71.5 percent of individual SEEG responses and 17.0 percent of whole pulses at 20 to 500 ms in eight sessions
Uses the prior joint score study to state the scope of this contribution. That study tested decoder scores after simulated contact loss in a different dataset. This study tests evoked waveform forecasts at complete stimulation pulses.
- usesQuadratic bounds certified 39.4 percent of later SEEG decisions for simultaneous constant contact gain changes of up to five percent in eight recordings
Its prediction sets under known contact loss informed the distinction between label coverage and decision stability. No fitted output was used.