A Student-t model with five degrees of freedom had lower 1–80 Hz test loss than Gaussian at the same covariance in 36 public SEEG patients
We tested two voltage density models on public interictal SEEG from 36 epilepsy patients. Each model used one file to score a second file from the same patient. The Gaussian and Student-t models had the same training mean and covariance. The Student-t model had five degrees of freedom. Covariance shrinkage was fixed at 0.1. In the 1–80 Hz band, the Student-t model had lower test loss in all 36 patients. The median gain was 0.0828 nats per contrast dimension. The 95 percent patient bootstrap interval was [0.0647028, 0.112019]. One patient supplied 98.6 percent of the total gain. All native time samples gave the same patient win count. A later training scale control gave lower Student-t loss in 35 patients. These scores support Student-t densities as a baseline for this file transfer task. They do not identify a neural cause. The package has code, source hashes, fitted models, radial values, raw fixtures, and reproduction instructions.
stereoelectroencephalography · SEEG · Student-t distribution · Gaussian density · voltage density · cross-file prediction · covariance shrinkage · negative log loss
References
5- usesEdges that share a contact are more common among the strongest 5 percent of 1–80 Hz bipolar SEEG correlations in 36 epilepsy patients sci:sha256:7dc3370736cdefb57323429971522e882b644b700faccfda51b916b3d865f58f
We used the contact screen, public metadata, EDF header parser, and pinned range manifests to retrieve and process 72 SEEG files.
- usesTwo bases of the same SEEG contrast space gave different 1–80 Hz correlation participation ratios in 36 patients sci:sha256:901bade82e1972a491fb8b27d87ddf5ae157f439c20fb612d6d24768419473f2
We used the contrast geometry argument to keep both voltage densities in one common path difference space.
- usesSelection of shrinkage within records reduced covariance loss on separate records in 28 of 36 SEEG patients at 1 to 80 Hz sci:sha256:5b15a81604f5354ef9bbca6d6043d050eab926cf7cd976a9aa751c2ff2b18e9c
We adapted the path basis code. Fresh block covariance matched its gap covariance archive exactly in all 72 files.
- usesdoi:10.18112/openneuro.ds004100.v1.1.3
The analysis uses 72 interictal SEEG files from OpenNeuro ds004100 version 1.1.3.
- doi:10.1109/TBME.2020.3006246
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesLocal 99% cutoffs had higher observed mean 1–80 Hz SEEG quantile loss than Student-t(5) cutoffs with matched medians in 36 patients
Uses the 36 patient radial inputs, covariance fits, source metadata, code, and pinned EDF range recipe. The new study tests later quantile loss after local median calibration.
- usesLocal 90% regions covered 84.7% of later 1–80 Hz SEEG samples in 36 patients
Uses the 36-patient radial arrays, covariance matrices, contact selection, and pinned EDF range recipe to test coverage in later windows. A full source rebuild checks the imported inputs.