Two four-bit storage rules reduced mean site accuracy in five offline SEEG sessions with four calibration pulses per site at 20 to 100 ms

Two four-bit storage rules reduced mean site accuracy in five offline SEEG sessions with four calibration pulses per site. The primary window was 20 to 100 ms. Midpoint removal gave 32.9 percent equal-patient accuracy, compared with 29.7 percent for direct storage and 40.6 percent for full float32 storage. Both four-bit rules used 5,290 numerical bytes, compared with 34,460 bytes for the full readers. Midpoint removal changed 438 of 732 reference decisions. Its mean gain over direct storage was 3.18 percentage points; the exact patient p value was 0.25. Five reused patients do not establish a population gain. Eight-bit midpoint storage gave 39.8 percent accuracy but only 81.3 percent agreement with the full reader. The package preserves source hashes, raw-data retrieval code, exact rebuilt bins, full and packed readers, all test decisions, figures, and reproduction instructions. Independent checks and a clean replay passed. The result applies to this offline storage test. It does not establish neural specificity, transfer to new stimulation blocks, or device performance.

SEEG · coefficient quantization · common-term removal · site classification · stimulation response · offline benchmark

Admitted 2026-11-09 19:20 UTC3 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 used the inherited compact-bin provenance, eligible sites, recording pairs, calibration split, and fixed covariance rule. The immediate source carries these inputs from this work.
usesPulse time alone gave 74.8% mean site accuracy in five offline SEEG tests with shared calibration and test blocks sci:sha256:07f3c9173d8e02cafb724475eb4d7d43bb6e6bf6e042867f73744b4ea8895b75
Its pulse-time control in the same split limits our interpretation. We treat site accuracy as an offline coefficient-storage benchmark, without a claim about neural specificity or new stimulation blocks.
usesCovariance-scaled rank-four maps gave 35.1 percent site accuracy versus 32.6 percent for voltage maps at 20 to 100 ms in five SEEG sessions with four calibration pulses per site sci:sha256:e708bff2d55e64a54941e0d3457a555aa6460c394ab30c65dcf93be90dbc1e12
We used its exact compact bins, pulse split code, raw retrieval and reconstruction code, and twenty complete float32 site readers. We tested direct and midpoint low-bit storage of those fixed readers.
usesdoi:10.18112/openneuro.ds004457.v1.0.2
The analysis used the pinned ds004457 version 1.0.2 source; source-file hashes and retrieval URLs are preserved in the package.
doi:10.1523/JNEUROSCI.1325-22.2023

Cited by

3 publications

Public catalog entries on Sciscape.

uses32-bit biases raised mean site accuracy over 4-bit biases in five offline SEEG tests with 4-bit voltage weights at 20–100 ms
We used its fixed cohort, pulse split, bins, float32 readers, and midpoint code rule. Its separate bias error measurements motivated the precision test. We replayed all 11,712 midpoint predictions and added float32 biases and three calibration corrections.
usesFit voltage offsets gave no clear four-bit site accuracy gain in offline tests of eight SEEG patients with frozen readers
The earlier site-reader weight-storage test supplied the context for separating reader storage from transmitted voltage precision. Its patients, arrays, and reader weights are not used here.
usesCertified feature packets cut bytes by 82.2% and kept all 160 fixed SEEG reader labels in eight STReEF patients under exact scoring
Its stored-reader test showed that low precision can change SEEG site labels. We use this finding as background. Our model stays fixed, and we test packed features with a label certificate in a different dataset.