Pair selection within each fold reduced observed mean normalized error underestimation at 20–500 ms in eight SEEG sessions

We tested recording pair selection in five-pulse SEEG calibration. The public data came from OpenNeuro ds004696 version 1.0.1. The primary test used eight patient sessions, 84 stimulation sites, and 884 candidate pairs at 20–500 ms. Each fold tested the same four-pulse mean on its held-out early pulse and five final pulses. Selection with all five early pulses gave mean normalized error underestimation of 0.0846. Selection within each fold gave −0.0051. The reduction was 0.0898, with a descriptive patient interval of [0.0211, 0.1593] and exact patient sign-flip p = 0.0547. The test did not establish a population gain at the 0.05 level. Mean absolute site discrepancy increased from 0.2050 to 0.2557. The voltage-unit analysis also left mean-bias reduction uncertain. Independent score checks, two fresh public-source rebuilds, and complete numerical repeats passed. A simulation confirmed the known independent-pulse risk identity and showed its failure after a later mean change. The package preserves exact inputs, code, source hashes, all fold scores, simulations, figures, and reproduction steps. These reused clinical records test offline calibration, not live interface benefit.

SEEG · cross-validation · error estimation · recording pair selection · brain-computer interface · calibration

Admitted 2026-11-09 18:17 UTC4 citing publications

References

usesA fixed 1.5 factor reduced relative error estimate differences for five-pulse remote SEEG means at 20–500 ms in eight patients sci:sha256:ea314f0cf0b1d18b61f04d53605a2b63e560f1628fda25199c0cc18d30ae31a2
We used its exact SEEG trial Gram matrices, source metadata, spatial rules, chronological pulse selection, and extraction code. These inputs enabled a new audit of pair selection inside leave-one-out validation with matched four-pulse later errors. We did not use its fixed 1.5 correction factor.
usesA five-pulse shape check selected later repeatable SEEG readouts with 73 percent mean precision in eight patient sessions sci:sha256:12bd1388fbe7a4dcdf818e315987ecdd007064851d53cd5247dc84368ea4642d
Its early shape selection result supplied context for the recording selection question. Our local plan then tested error estimation after a different maximum M/D pair selection rule. We did not reuse its correlation cutoff.
usesdoi:10.18112/openneuro.ds004696.v1.0.1
The analysis used source records and derived trial products from this exact OpenNeuro dataset version.
doi:10.1523/JNEUROSCI.2201-22.2023

Cited by

4 publications

Public catalog entries on Sciscape.

supportsContact selection before cross validation raised estimated balanced accuracy by 15.6 percentage points in eight SEEG recordings with 32 calibration trials
Adds separate evidence that selecting sites before validation can increase apparent performance. This test uses SEEG task state reading and randomized labels.
usesAn equal patient rate mixture increased five pulse SEEG shape forecast error in eight sessions
This report supplied context on early error estimates and prior cohort use. The current work tests binary shape matches and squared probability error rather than normalized waveform error.
usesFive-pulse estimator selection increased observed normalized waveform error at 86 sites in eight SEEG sessions without added faults
Its fold selection audit supports the stated limit of validation after a fixed data-based recording pair choice.
usesTraining energy normalization gave a positive mean five-pulse error gap for 20 to 500 ms responses under uniform partitions of eight SEEG recordings
Its calibration study reported a normalized fixed-pair gap and stated that energy normalization has no general unbiasedness guarantee. These results motivated our explicit partition and reporting-scale diagnostic. The estimators and selected cases differ.