Background voltage prediction reduced mean centered SEEG error at 20–100 ms by 3.89 percent compared with four-pulse templates at 94 later sites in five sessions
We tested a model that predicts voltage after an electrical pulse from voltage before the pulse. The data were public SEEG from OpenNeuro ds004457, version 1.0.2. The shared linear maps used only background voltage from earlier sites. Each later site had a template from its first four pulses. The model adjusted this template with the test pulse history and a calibration correction. The last four pulses were scored. Across 94 sites, 376 test pulses, and five sessions, mean centered 20 to 100 ms waveform error was 3.89 percent lower. Patients had equal weight. The 95 percent patient bootstrap interval was 0.31 to 9.30 percent. The exact two-sided patient sign-flip reference gave p = 0.125. A matched test before stimulation gave 9.10 percent mean gain. The gain was not unique to a pulse window. This result does not show prediction of a neural change caused by stimulation. The cohort and contacts came from earlier work, so this is an exploratory study. Rebuilding the inputs from raw data gave exact matches in all five sessions. The package preserves inputs, fitted models, forecasts, source hashes, code, figures, and reproduction instructions.
stereo-EEG · intracranial EEG · voltage forecasting · electrical stimulation · ridge regression · brain-computer interfaces
References
7- usesPooled site sets included the correct site for 93.6 percent of later SEEG pulses within five sessions at 20 to 100 ms sci:sha256:16a0821e6ea5302515dea9c46cffb63ec910cd0b438ee578fb5d0ecad4563dc7
Used its native MEF retrieval and disjoint 5 ms extraction recipe through the immediate source. We verified 110 source files by size and SHA-256 and rebuilt all five sessions. The package preserves the source manifest and MIT notice.
- usesFour signed pre-pulse modes per pair raised mean relative 20–100 ms SEEG forecast error by 19.2% at 94 later sites in five sessions with ridge strength 0.1 sci:sha256:cff1a65f787100749ac1ffac63d759ad314b24838dfcad17b00541495cbdf570
Used its compact contact voltage inputs, event tables, fixed contact settings, source metadata, loader, and retrieval recipe. Fresh raw extraction matched all five sessions exactly. The new model uses only background voltage for shared weights. Its baseline differs from the earlier model, so their reported gains are not a matched numerical comparison.
- usesAll earlier pulses gave lower mean SEEG forecast error than four recent pulses in five recordings at 20–500 ms sci:sha256:7572e2f1a54f04ee36ae5756aff12b06f8ed5349904c849467a91c2e6236dbe9
Used its 5 ms voltage benchmark, fixed recording contacts, and pinned source subset through the immediate source. These inputs define the tested recording pairs. All five raw rebuilds matched the preserved voltage arrays and event tables.
- usesA fixed pre-pulse voltage model had a mean 3.8% error increase at later sites in five SEEG sessions sci:sha256:75c9b7d9135d45ecf7f9a8eb06ff6798fe1e048c1713980e8ef4f61e237bf529
Used its source quality loader and strict earlier-to-later site split through the immediate source. The saved split matched the full split in this study. We retained the main loader and documented its stimulation-contact status limit.
- usesdoi:10.18112/openneuro.ds004457.v1.0.2
Source SEEG voltage and stimulation events for the analyzed OpenNeuro ds004457 version 1.0.2 dataset.
- supportsdoi:10.1523/JNEUROSCI.1325-22.2023
Source study describing the ventral temporal cortex stimulation dataset used here.
- doi:10.1111/epi.17528
Related motivation on pre-stimulation phase and response variation.
Cited by
4 publicationsPublic catalog entries on Sciscape.
- usesBackground subtraction gave uncertain site log loss gains with three template pulses in five SEEG sessions
Used its compact public voltage data, fixed readout pairs, chronological site split, source hashes, retrieval and extraction code, and background map design. The fitted maps match its values. The new test measures separately calibrated site probabilities with three template pulses.
- usesA background model lowered mean voltage error with an early reference and increased mean shape error at 20 to 100 ms after five fit pulses at 86 sites in eight HAPwave SEEG sessions
Its background-only linear forecast and calibration-mean correction provide the related method. This study tests a site-specific five-trial fit in HAPwave with voltage-level and shape checks. No source code was copied from that study.
- Background subtraction reduced mean SEEG site accuracy at 20 to 100 ms with four fit pulses per site in five offline sessions
[Legacy relation: reproduces] The five fitted main background maps matched its archived parameters exactly. The same 94 sites and 376 pulses gave its 3.89 percent mean waveform MSE reduction. This is a computation check on reused data, not an independent cohort result.
- usesBackground subtraction reduced mean SEEG site accuracy at 20 to 100 ms with four fit pulses per site in five offline sessions
We reused its input loader and background map design, including earlier sites, translated windows, and fixed ridge setting. We tested whether its waveform forecast gain transfers to stimulation site reading.
- usesLinear maps fitted to pulse responses had more 20 to 100 ms SEEG forecast error than background maps in five selected sessions at ridge strength 0.1
Its background forecast result motivated the matched response-versus-background question. We used its idea of a four-pulse site mean plus a centered history correction. Our windows, fit sample budget and time split differ, so this is not a replication of its gain.