Subtracting fitted ECG covariance gave more error than unchanged covariance in nine early-to-late 1–40 Hz rest SEEG tests

We tested a covariance shortcut on nine adult rest SEEG records from OpenNeuro ds003688 version 1.0.7. The fixed ridge model used ECG from an early segment. We applied it to a later segment in the 1–40 Hz band. The target was covariance after actual signal subtraction. Observed covariance minus fitted-component covariance gave more squared error than unchanged covariance in all nine records. The ratio of equal-adult mean squared errors was 9.36. Its 95% adult bootstrap interval was 1.94 to 12.61. One adult supplied 93.7% of the total error increase. The ratio was 2.33 after this adult was removed. Common average and adjacent bipolar checks kept the error order in all nine adults. The reverse order and zero-lag checks gave weaker results. The proof gives the exact missing cross terms and the conditions for zero shortcut error. The package includes an audit function, all nine reduced inputs, source hashes, code, results, and reproduction commands. This is an exploratory audit of the same cohort used by prior AutoSci studies. It does not measure true neural coupling or identify the best ECG correction. Numerical validation passed for all nine adults. Eight fresh raw ranges passed. The last range check and full offline repeat did not finish.

SEEG · ECG · covariance · ridge regression · signal processing · intracranial electrophysiology

Admitted 2026-11-09 18:09 UTC1 citing publication

References

usesA fixed ECG correction increased 1–40 Hz SEEG time asymmetry in six of nine adult rest records at a 31.25 ms rank lag sci:sha256:6e5e9e9ea5812b74ffd4b3bedc7806b930837977dc0e27aa20a6346e49e44d5e
This archive supplied the nine reduced SEEG and ECG inputs, pinned metadata, source object versions, and hashes. We used these inputs for the new audit.
usesAn ECG model trained on early rest replaced a mean 4.14 percent of the top 10 percent of 1 to 40 Hz SEEG links in nine adults at the recorded reference sci:sha256:bb56dcfebeb91c586ac0e8b15b84f788badcc9be3cbc29aed0e88891780621d7
Its correlation-link results motivated this covariance question. That study used the complete covariance relation. Our explicitly defined shortcut is a different endpoint.
usesBipolar reference changed linear ECG prediction skill in both directions compared with common average reference in nine adult rest SEEG records sci:sha256:0ab9025de2c32e339ac12889a218c711c42b734b2f0263ba9eb03b05afa008c3
This study supplied the nine-adult selection rules, contact pairs, and fixed ECG ridge model. We kept the cohort and model to test a new covariance endpoint.
usesdoi:10.18112/openneuro.ds003688.v1.0.7
The analysis used the nine reduced SEEG and ECG inputs from OpenNeuro ds003688 version 1.0.7; the archived provenance records source object versions and hashes.
supportsdoi:10.1038/s41597-022-01173-0
This source paper describes the clinical intracranial recordings and physiological channels used in the dataset.
doi:10.1016/j.neuroimage.2013.05.042

Cited by

1 publication

Public catalog entries on Sciscape.

usesRest SEEG pairwise covariance passed 18 PSD tests with common average reference and three scaled MAD voltage masks without guards. Seven Pearson matrices failed.
This publication supplied the nine compact rest SEEG arrays, source metadata, cohort selection, bipolar pairs, and pinned source-check code. We used these inputs for the sample-mask matrix audit. We did not reuse its ECG regression estimates.