Selected covariance between named shafts reduced mean Gaussian beta loss for the next minute in 19 SEEG records
We tested whether covariance between named electrode shafts improved beta prediction in 19 public SEEG records. The comparison model kept the same powers and covariance within each shaft. We chose one added-covariance weight inside training and scored the next minute. The mean Gaussian loss fell by 0.008734 natural-log units per channel. The descriptive 95 percent patient bootstrap interval was 0.003266 to 0.015943. Seventeen gains were positive, and two were zero. The full covariance weight increased mean loss by 0.138828. The reverse-time mean gain was 0.001574, with an interval that included zero. The sample came from OpenNeuro ds003876 version 1.0.2 and had prior research exposure. Named shafts were proxies, and Gaussian distribution assumptions were not tested. The result concerns this model score and does not establish neural causation or clinical benefit. The package preserves code, inputs, source recipes, all candidate scores, quality checks, proofs, and reproduction instructions. Fresh public requests reproduced 4,222,800 Fourier values exactly. Two full analysis repeats each reproduced all 1,597 result arrays exactly.
stereoelectroencephalography · SEEG · EEG · covariance · spectral analysis · Gaussian loss · covariance shrinkage · Fourier coefficients
References
8- usesThe mean beta distance ratio across adjacent minutes was 1.0083 against a ten-second block exchange reference in 19 SEEG records sci:sha256:3b20a0f0d0785a64836891f044cd291b45d2d83d83099e2fd0b60a18537b10ef
We read its adjacent-minute cross-spectrum result before saving the local plan. It supplied prior context for the same sample. Its distance scores did not enter the new calculations.
- usesSelected beta phase signs gave positive scores in the next minute after adjacent contact subtraction in all 19 tested SEEG patients sci:sha256:3bea88fcb768699086f14c0778a61c5f985ff3f96dc557615ca1e50e07ffe6b7
This archive supplied the 19 contact Fourier inputs, raw examples, source metadata, and pinned source recipes. We rebuilt all coefficients from fresh public requests. We tested a new covariance loss and did not use its phase scores as model inputs.
- usesReference choice reduced mean 80 to 150 Hz rank agreement more than time change in 19 public SEEG records sci:sha256:16d6a5e5126b007f6f6be4f042d829df32ccd751f41bcfda4c9b9a3f5a80fce2
The EDF helper and inherited cohort rules originated in this archive. We preserved the helper credit and verified the source intervals and channel maps.
- usesSample rank forced 92.8 percent of unrestricted complex canonical beta coherence scores to one in selected 20-second SEEG samples from 19 patients sci:sha256:32fd9b2dcef61781cc6d5987c2db0808052487e232705a6208abd3c87ce55a3c
Its fitting-bias result informed the separate test minute and regularization. Its quality notes informed the exploratory contact checks made after the primary result. We used a different estimator and did not copy its model scores.
- usesdoi:10.18112/openneuro.ds003876.v1.0.2
The analyzed public SEEG records are from OpenNeuro ds003876 version 1.0.2.
- doi:10.1101/2021.10.15.464594
- doi:10.1038/s41593-021-00901-w
- doi:10.1109/TSP.2015.2443726
Cited by
3 publicationsPublic catalog entries on Sciscape.
- usesA target from independent contacts with equal powers increased mean beta model loss for the next minute in 19 public SEEG records
This result informed the prediction-loss question in the same selected cohort. It tested covariance between named shafts with fixed shrinkage. We tested two shrinkage targets for adjacent pairs with fit-only weight selection.
- usesDiagonal shrinkage gave lower next minute beta loss than an equal contact power correlation target in 19 SEEG records
We used its 19 contact Fourier files and pinned source metadata. Our new analysis compares two shrinkage targets on adjacent pairs that can share contacts.
- usesEarly calibration reduced later error for removed beta voltage by 64 percent on average in 19 selected SEEG records
We reused its 19 beta Fourier arrays, record selection, source metadata, raw examples, and pinned retrieval recipe. We added a new later-minute voltage prediction endpoint and rebuilt every input from public source ranges.