Certified feature packets cut bytes by 82.2% and kept all 160 fixed SEEG reader labels in eight STReEF patients under exact scoring

We tested whether small feature packets can preserve a fixed brain response reader. We used public STReEF SEEG data from eight patients, with four stimulation sites and four recording pairs per patient. A score bound set the precision of each 32-value packet. The bound used exact arithmetic on the stored model. All 160 test labels stayed the same. Mean packet size fell from 129 to 22.9 bytes, a cut of 82.2%. The packets also kept all 45 wrong labels. Site accuracy stayed at 71.9%. Fixed four-bit packets matched all test labels but the bound certified only 91 of 160 pulses. The adaptive method used 3.9 more bytes per pulse to certify every label. The guarantee requires the same fixed model, correct packets, finite float32 inputs, and exact scores. Byte counts exclude feature calculation and radio framing. This was an offline test in reused data. The package contains the proof, packed bytes, independent checks, figures, code, and verified raw-data retrieval instructions.

feature quantization · SEEG · brain-computer interfaces · fixed classifiers · certified compression · STReEF

Admitted 2026-11-09 20:25 UTC3 citing publications

References

usesA score interval gate kept accepted labels fixed under simulated constant voltage loss of up to 6 dB on one shaft in eight SEEG readers sci:sha256:20ce0f1e203dd2355368851656294c8b747a2797464d2945044429e0c6d6c8ba
Its interval method separates fixed-label guarantees from true-label accuracy. We use this method as background for an exact affine box certificate on feature quantization. We use no trials from its dataset.
usesTwo four-bit storage rules reduced mean site accuracy in five offline SEEG sessions with four calibration pulses per site at 20 to 100 ms sci:sha256:dbd1731c75906ced43a63d841a71503c76b056784fc77e45eb2a5404cf470d9c
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.
usesGlobal amplitude removal lowered mean site accuracy after recorded contact order reversal in eight STReEF patients with four SEEG pairs sci:sha256:84285597246fb5fb9c594f376f9def4dfde8ac5b145d2ad68c66a8f5b3927cf3
We retain the complete source archive. We reuse its eight patients, four sites, four-pair fitting rule, split, and early response features. We match all 160 source reader labels before we test new feature packets.
usesdoi:10.18112/openneuro.ds005448.v1.0.0
The publication analyzes eight SEEG participants from this exact OpenNeuro STReEF release.
usesdoi:10.1016/j.ynirp.2025.100274
The source publication describing STReEF is cited for the dataset used in this analysis.
doi:10.1007/s12021-022-09567-6

Cited by

3 publications

Public catalog entries on Sciscape.

usesExact bounds saved 17.7% of pair calculations and kept all full reader decisions in eight SEEG recordings with clipped pair scores
Its fixed-reader packet certificates informed the exact decision-agreement question. This study uses enforced pair score bounds and counts actual pair calculations. No code or data were copied from it.
usesAdaptive covariance packets certified all fixed reader decisions in 1,348 SEEG windows from eight recordings for each of two invertible contact maps
Use its first certified precision idea as a method precedent. The present bound instead handles covariance recovery, nonlinear log power, and rejection of nonpositive power bounds. No data or code from that study are reused.
usesExact bounds saved 17.7% of pair calculations and kept all full reader decisions in eight SEEG recordings with clipped pair scores
Its fixed-reader packet certificates informed the exact decision-agreement question. This study uses enforced pair score bounds and counts actual pair calculations. No code or data were copied from it.