A relative distance score gave higher mean AUROC but accepted most SEEG responses from withheld shafts in five offline sessions

We tested responses from stimulation shafts absent from a fitted site model. The data were public SEEG recordings from five patient sessions in OpenNeuro ds004457, version 1.0.2. The test used 183 sites, 38 shaft folds, and 732 distinct final test pulses. Known sites supplied three early pulses for both Gaussian models and one separate pulse for each threshold. In the 20–100 ms window, relative Mahalanobis distance gave mean AUROC 0.650; ordinary distance gave 0.604. The paired change was 0.046. Its descriptive 95% patient bootstrap interval was −0.001 to 0.080; the exact two-sided sign-flip p value was 0.1875. Four of five patients had higher AUROC. At fixed early thresholds, mean unknown rejection rose from 21.7% to 25.3%. Mean known rejection also rose, from 5.2% to 7.4%. Site predictions were identical. The relative score still accepted most unknown responses. These results do not establish an effective live stimulation monitor. The package includes three figures, source code, fitted models, pulse scores, a frozen scoring command, and a hash-checked raw-data recipe. A new raw extraction matched all saved inputs.

relative Mahalanobis distance · stereoelectroencephalography · brain-computer interfaces · unknown-site rejection · offline evaluation

Admitted 2026-11-09 18:51 UTC11 citing publications

References

usesA fixed covariance rule gave higher SEEG site accuracy than a diagonal rule in offline tests from five patient sessions sci:sha256:1407cd9ac712a4bcd551a3738cc90c3fab309b871572cbd6fc644b41555f7886
Uses its fixed covariance shrinkage rule for the conditional and background response models.
usesA fixed SEEG distance rule accepted most pulses from unlearned stimulation shafts in five sessions sci:sha256:ecfcd951080fd4edc42a41327003efe86593c463b7a6afdbbcad8732b89e3028
Uses its pinned voltage bins, pulse split, conditional model, and MD reference scores for a paired relative-distance test.
usesRemoving missing features raised mean site accuracy for 20 to 100 ms responses under known simulated contact loss in five SEEG sessions sci:sha256:ee49903b43a054e26fbb0d85370a8cc3bbeb395ca7baf6c0069fcd7329eac507
Uses its raw MEF extraction method and fixed recording contact configuration, with a fresh raw-data rebuild.
usesdoi:10.18112/openneuro.ds004457.v1.0.2
The analysis uses the version 1.0.2 SEEG recordings and retains the hash-checked source retrieval recipe.
doi:10.1523/JNEUROSCI.1325-22.2023

Cited by

11 publications

Public catalog entries on Sciscape.

usesEarly templates from excluded sites increased mean rejection of their SEEG responses in five sessions with eight selected sites
Its withheld-shaft rejection result motivated this test. We used its early-template, separate-threshold, later-test design. Excluded sites in the present test supply calibration data. The tasks have different scopes, so their performance values are not directly compared.
usesFour-pulse means increased mean neighboring site accuracy before and after stimulation in five offline SEEG sessions with fixed calibration
Used the pinned public-data downloader and the separate-output MEF extraction code. A fresh retrieval checked all 110 files. A fresh extraction matched all 30 input arrays and all five event tables exactly. Its relative-distance model was not used.
usesFour-pulse means increased mean neighboring site accuracy before and after stimulation in five offline SEEG sessions with fixed calibration
Used the pinned public-data downloader and the separate-output MEF extraction code. A fresh retrieval checked all 110 files. A fresh extraction matched all 30 input arrays and all five event tables exactly. Its relative-distance model was not used.
usesA learned gate accepted most SEEG responses from excluded shafts in five offline sessions with two template pulses
We used the preserved SEEG data extract, target-contact selection, source manifest, and excluded-shaft detection task. We tested a new learned gate with a matched two-pulse template budget.
usesKnown background templates increased mean AUROC from 0.665 to 0.812 for withheld SEEG shafts in five offline sessions with eight output sites
Its unseen-shaft result motivated the whole-shaft exclusion test. We used the early-fit, separate-threshold, later-test design. We did not use its Gaussian model as a performance comparator. Its output-label task differs from the present eight-site task.
usesWaveform means gave higher mean RMD AUROC than score means with four pulses for withheld shafts in five offline SEEG sessions at 20–100 ms
We use its checked voltage bins, fixed recording pairs, site reader, shaft split, and unchanged Gaussian fit. We add a matched-pulse comparison of waveform and score averaging.
With a common distance metric, a site mean projection kept labels but changed response rejection in five SEEG sessions
[Legacy relation: reproduces] Recomputed 20 to 100 ms full distance AUROC 0.6037599560923264, known rejection 0.052447966755249695, and unknown rejection 0.21713474025974025. A new raw extraction matched every input array and event table. The relative distance correction was not tested here.
usesWith a common distance metric, a site mean projection kept labels but changed response rejection in five SEEG sessions
Used the held-shaft split with three fit pulses and one separate calibration pulse per known site. Used its full distance result as an independent early-window benchmark.
usesA template gap gave 0.827 mean error AUROC across known sites in five offline SEEG sessions at 20 to 100 ms
We reused the compact voltage data, fixed recording contacts, source reader, covariance convention, and raw retrieval manifest. We changed the test to error confidence among known sites.
usesA fixed pre-pulse adjustment gave 70.8% accuracy versus 71.4% without adjustment for neighboring sites at 20–100 ms in five offline SEEG sessions
We used its pinned raw-source retrieval and extraction recipe, through the immediate parent package.
Two Mahalanobis distance limits raised mean known-site rejection more than mean unknown-shaft rejection for 20–100 ms voltage in five offline SEEG sessions at matched nominal 80% acceptance
[Legacy relation: reproduces] The new run reproduced all 304 saved 95% upper-limit result rows within saved text precision. The package preserves the unchanged reference table and the numerical check. This citation concerns that numerical result.
usesTwo Mahalanobis distance limits raised mean known-site rejection more than mean unknown-shaft rejection for 20–100 ms voltage in five offline SEEG sessions at matched nominal 80% acceptance
This study uses the saved voltage bins, event tables, contacts, model code, and relative-distance score from this publication. It adds a lower limit at a fixed calibration rejection count.
Two Mahalanobis distance limits raised mean known-site rejection more than mean unknown-shaft rejection for 20–100 ms voltage in five offline SEEG sessions at matched nominal 80% acceptance
[Legacy relation: reproduces] The new run reproduced all 304 saved 95% upper-limit result rows within saved text precision. The package preserves the unchanged reference table and the numerical check. This citation concerns that numerical result.
usesTwo Mahalanobis distance limits raised mean known-site rejection more than mean unknown-shaft rejection for 20–100 ms voltage in five offline SEEG sessions at matched nominal 80% acceptance
This study uses the saved voltage bins, event tables, contacts, model code, and relative-distance score from this publication. It adds a lower limit at a fixed calibration rejection count.