A contact mixture raised mean 20 to 100 ms site accuracy under hidden simulated contact errors at ten times the fitted voltage scale in five SEEG sessions
A reader must use noisy brain voltage without a reliable fault label. We tested a fixed Gaussian mixture that uses the signs of adjacent-contact differences. The source was OpenNeuro ds004457 version 1.0.2. Five patient sessions gave 183 stimulation sites and 732 distinct test pulses. We added a hidden error to one recording contact at a time. In the primary test, we set the error standard deviation before time centering to ten times the fitted voltage scale. The scale used earlier pulses at each site. Mean 20 to 100 ms site accuracy rose from 19.0% with an ordinary Gaussian reader to 34.7% with the contact mixture. Each patient had the same weight. The gain was 15.8 percentage points. The exact two-sided patient sign-flip test gave p = 0.0625. Clean accuracy changed from 40.6% to 40.5%. A mixture that assigned the error to one recording pair gave 30.1% accuracy. A shared physical contact can affect two pairs with opposite signs. The contact mixture preserved this relation. A delayed window and smooth added errors also gave positive mean gains. The package contains the fixed protocol, voltage bins, models, random errors, all decisions, figures, and source code. All five voltage archives matched a fresh raw extraction. An analysis replay matched all saved model and error arrays. This exploratory test concerns simulated errors and known stimulation sites. It does not establish performance with real hardware faults or a live brain interface.
stereoelectroencephalography · single-pulse electrical stimulation · brain-computer interface · contact geometry · Gaussian mixture · simulated contact error · site classification · signal processing
References
4- usesA fixed covariance rule gave higher SEEG site accuracy than a diagonal rule in offline tests from five patient sessions sci:sha256:1407cd9ac712a4bcd551a3738cc90c3fab309b871572cbd6fc644b41555f7886
We used its fixed half-covariance site reader and the same first-four fit and last-four test allocation through publication 85133706-92de-4e53-bdc8-810fad6939b6. The new CSV and NumPy reader matched the inherited cohort, features, means, and covariance for all 15 subject-window models. We read this report and preserved its metadata. Its cohort is reused, not counted as new evidence.
- usesCovariance projection gave 29.1% mean site accuracy after known simulated contact loss in five SEEG sessions at 20 to 100 ms sci:sha256:4129d22937b1254193093910d60029692219399b8b4fca380b01959c3aa2b5ed
We reused its checked 5 ms voltage bins, source metadata, raw retrieval and extraction code, fixed inclusion rules, and base covariance fit. Its known contact-loss test motivated this new hidden additive-error test. We rebuilt all five inputs from raw data and retained the same patients, so this is not an independent patient replication.
- usesdoi:10.18112/openneuro.ds004457.v1.0.2
Source dataset for the five SEEG sessions analyzed in this publication.
- usesdoi:10.1523/JNEUROSCI.1325-22.2023
Scientific source for the dataset and its collection methods.
Cited by
3 publicationsPublic catalog entries on Sciscape.
- usesCentered median templates raised mean site accuracy with one simulated wrong label per four fit pulses and lowered clean accuracy in five SEEG sessions at 20 to 100 ms
We reused its checked ds004457 voltage bins, fixed recording contacts, site eligibility, source manifest, and retrieval and extraction utilities. We tested new fit-label permutations, not its simulated contact voltage errors.
- usesCycle checks missed simulated contact errors that changed SEEG site decisions in five offline sessions
Its shared-contact simulated-error study defines the relevant signal-chain fault model. Our study keeps the ordinary reader fixed and measures targeted minimum error size instead of fitting or comparing mixture readers.
- usesLater 20 to 100 ms SEEG variance outside a shape fitted from four pulses was 1.20 times the variance before pulses at 138 sites in five patients
We used its checked voltage bins, fixed recording target list, source metadata, and raw retrieval and extraction code. We rebuilt all five voltage archives from the pinned source objects. We kept the same patients and used a new site rule and variance outcome. This is not a new patient sample relative to that input publication.