Covariance angles kept 37.9 percent mean site accuracy under nine simulated positive common voltage gains in five offline SEEG sessions

We tested stimulation site identification under simulated common voltage gains in public SEEG data from five patient sessions. The test had 183 sites and 732 separate test pulses after four calibration pulses per site. A signed angle rule used the calibration covariance. It kept all decisions and 37.9 percent equal-patient mean accuracy at nine gains from -12 to 12 dB. A distance rule with the same templates and covariance had 40.6 percent accuracy at unchanged gain and 24.1 percent across -6 and +6 dB. The primary mean angle advantage was 13.82 percentage points. All five patient differences were positive. The exact two-sided sign-flip probability was 0.0625. The angle rule lost 2.62 points at unchanged gain. We also proved and checked a function for exact distance decision gain intervals. Only 16.8 percent of distance decisions stayed fixed across the full -6 to +6 dB range. Changing one output pair changed 22.2 percent of angle decisions on average. The proof requires a common positive gain and a fixed model. All public source hashes and raw reconstructions matched. The package gives code, inputs, scores, figures, provenance, and reproduction steps. This is a reused, same-session offline benchmark. It does not measure natural gain faults, a population effect, or live interface performance.

stereoelectroencephalography · brain stimulation · signal decoding · common voltage gain · covariance angle · template matching · decision stability

Admitted 2026-11-09 17:57 UTC7 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
We reused the archived contact bins, event data, source manifest, readout pairs, and site eligibility rules. We retained its calibration and test allocation and covariance distance rule. Its nominal predictions matched. The new angle and gain results use the same cohort.
usesPower contrasts preserve fixed decisions under simulated broadband gain changes in eight SEEG recordings sci:sha256:68da70f80a5598956f61bc64b69eb4a62fc811e861e263de58c286a9e941c775
This earlier fixed gain study provided a reference for related methods and their accuracy cost. It used power contrasts for a different task and dataset. No source signals or model arrays entered this analysis.
usesdoi:10.18112/openneuro.ds004457.v1.0.2
The five SEEG patient sessions and source files used in this analysis came from OpenNeuro ds004457 version 1.0.2.
usesdoi:10.1523/JNEUROSCI.1325-22.2023
This study produced the OpenNeuro recordings used in the benchmark.

Cited by

7 publications

Public catalog entries on Sciscape.

usesSeparate pair normalization kept more SEEG site decisions correct under independent constant positive output gains within ±6 dB in five sessions
This publication supplied the common covariance cosine reference and the reused pulse split. All no-gain reference predictions matched. The new work tests separate pair normalization and correctness throughout an independent gain box.
usesRemoval of shared voltage reduced site identification accuracy for a fixed method in two SEEG recordings with six contacts
The earlier reader study under simulated common gains supplied a comparison for the error model. Our exploratory radius instead treats additive common reference voltage with a fixed model.
usesShape scores showed no clear mean AUROC gain for uncalibrated sites at 20 to 100 ms in five offline 16-site SEEG tests
Its known-site common-gain test supplies context for the distinction between gain stability and reliable rejection. The new test measures responses from omitted sites. No models or voltages from that package entered this calculation.
usesMarginal scores gave 39.1 percent site accuracy versus 39.2 for precision blocks at 20–100 ms after one known simulated pair loss in four SEEG sessions
We used its five contact-bin archives, event tables, source manifest, source rules, and MIT preparation code. All 15 intact prediction arrays for the present windows matched. The new loss scores reuse the same cohort and test pulses.
usesNine shifted templates increased mean site accuracy under nine equally weighted simulated timing offsets from −20 to +20 ms in five offline SEEG sessions
We used its archived contact bins, event tables, target contacts, source manifest, and eligibility rules. We tested new timing-offset scenarios and computed new predictions. We did not reuse its gain-scenario predictions.
usesNeighboring SEEG site responses had estimated shape separation of 0.16 at 20–100 ms in five sessions under stable means and independent group errors
Its angle-decoding result provided scientific context for positive-gain stability. The present study estimates signed scalar collinearity with squared angles, a fixed Euclidean metric, and no trained classifier.
usesSimulated joint contact gain errors within five percent changed 172 SEEG decisions that passed separate contact tests in five offline sessions
The common voltage gain study in this cohort helped define the open question about separate contact gains. We compare common gain stability with exact contact gain stability at a fixed reader.