A check of two scores gave 27 false guarantees for 10 µV contact errors in 20–30 ms bins in five offline SEEG readers
We tested voltage guarantees for five fixed SEEG readers with public OpenNeuro ds004457 data. We reused a fixed 183-site cohort and tested 732 pulses. Errors could affect every recording contact in the 20–30 ms bins, with an assumed absolute cap of 10 µV. A check of the two largest site scores passed 170 decisions. It gave 27 false guarantees. A stored error below the cap changed each of those 27 labels. The exact check of all rival sites passed 143 decisions: 126 correct and 17 wrong. A conservative pair check passed 132. A bound can preserve a wrong label. The package gives a proof, code, fitted readers, source hashes, raw retrieval instructions, test results, and four figures. Raw extraction matched the inputs exactly. Independent linear programs checked 1,464 boundaries. This is a new error audit of a reused cohort. The cap is a stress assumption. The study does not establish an online benefit or performance in a new patient.
stereoelectroencephalography · robustness certificate · contact voltage error · linear classifier · brain-computer interface
References
2- usesA voltage bound for one contact certified five more correct labels in eight offline SEEG readers at a fit voltage benchmark with fixed references sci:sha256:e9e327272c436b8db922084ff96586b00de8eec931a1e5517904ac6e33c1e7ee
This prior study used the physical contact to pair map for a single-contact RMS bound in another SEEG cohort. Its comparison informed our exact contact and conservative pair checks. Our main result tests omitted rival sites under a cap on all short contact-bin errors.
- usesOne shared temporal noise matrix gave lower mean SEEG site accuracy than full covariance with two fit pulses per site in offline tests of five sessions sci:sha256:80660247e5d8e43c18f311062884ad6ed208a80d726410b122af21a5b954b4c0
This publication supplied the fixed five-patient cohort, source contact selection, extraction code, and full covariance reader. We rebuilt its inputs from raw files and matched all 732 nominal predictions. We then tested the fixed readers under a new contact error model.
Cited by
1 publicationPublic catalog entries on Sciscape.
- supportsMean AUC lower bounds exceeded 0.5 for site errors capped at 5% of training response SD in 20 iEEG patients and 20 fMRI adults
It shows why a robustness certificate must cover the full stated error model in a different SEEG reader. We do not reuse its data or code.