Selected SEEG pair deviations reduced mean error in log sound amplitude on 100 later passages in nine people with a 400 ms output delay

We tested whether selected SEEG pair power adds useful information beyond mean log power across all eligible pairs. We used public OpenNeuro ds004703 version 1.1.0. Six early passages per person set pair selection and ridge penalties. The test used 100 later passages in nine people, with a fixed 400 ms output delay. The pooled reader used ten columns. The added-pair reader used 90 columns, including deviations from eight selected pairs. Mean error skill rose from 0.0025 to 0.1427. Skill is one minus test mean squared error divided by the error of the frozen training target mean. The mean paired gain was 0.1402; its descriptive 95% person bootstrap interval was 0.0604 to 0.2202. Seven people improved and two had more error. A later exploratory control with 32 random pair sets per person had mean skill 0.0205. Fresh retrieval and independent spectra matched 4.70 million power values and all 48 sound target files. Independent model checks and complete replays passed. The package preserves features, models, predictions, four figures, code, source identities, and retrieval instructions. This is a reanalysis of test data used in prior work. Both main readers need all eligible pairs to compute the mean. The result supports preservation of pair detail in this same-session heard-speech reader. It does not show performance for intended speech, new sessions, fewer acquired contacts, or brain stimulation.

stereoelectroencephalography · SEEG · brain-computer interface · sound amplitude · spatial features · ridge regression

Admitted 2026-11-09 19:35 UTC9 citing publications

References

usesFixed SEEG readers had higher mean sound amplitude correlation at 400 ms than at zero delay on 100 unused later passage records in nine people sci:sha256:3c8b962f91454c5d1b32469fb10a6ecb9534841b65f6b7ab946bc56a062e2126
The study uses corrected-clock power features, the six early training passages per person, all 100 eligible later passage records, and exact source identities from this package. It fits new pooled and pair-deviation readers. The later records are a reused test set.
usesTen offline SEEG readers gave mean correlations of 0.214 at zero added delay and 0.302 at 400 ms for heard sound amplitude sci:sha256:0363ee7f6d1912a2c732d3938bfa9b204378683c1ded14e573a003ec7c3ef2fc
The study uses the early passage rule, corrected source-clock method, power extraction rule, and 400 ms delay established in this method package. The new study tests spatial feature information with training-only whole-passage validation and new reader fits.
usesdoi:10.18112/openneuro.ds004703.v1.1.0
The study used OpenNeuro ds004703 version 1.1.0 SEEG and sound data. Its dataset description states CC0, while its README states noncommercial and no-identification restrictions; the conflict is unresolved and source data are omitted from this package.

Cited by

9 publications

Public catalog entries on Sciscape.

usesTraining for one known feature group loss increased mean worst-case skill and decreased mean clean skill in nine SEEG sound readers
This study uses the compact SEEG power arrays, sound targets, passage lists, and source identities. It fits eight raw pair readers and tests exact loss training. The 100 later passage records are reused.
usesTime contrasts raised mean sound amplitude skill by 3.54 points over band ratios on 100 later SEEG passages in nine people at 400 ms delay
This publication supplied the two-band SEEG features, sound targets, early training split, later test split, source metadata, and checked extraction recipe. The present study fits new nested gain-protected readers on those inputs.
usesBands from four calibration passages per person covered 85 of 100 later sound traces with 117 points in nine SEEG recordings
This study uses the compact adjacent-pair power features and target extraction inherited from this work. It preserves the source notices and checks the features against fresh public source ranges.
usesCalibration at nine offsets showed no clear mean sound readout gain in nine SEEG records with simulated offsets from −400 to +400 ms
This study uses its compact power arrays, sound targets, source pins, and early-versus-later passage split. The test records were reused. This study adds nine simulated offsets and a new calibration comparison.
usesA shared time filter changed mean SEEG sound amplitude skill by −0.00018 in nine listeners at a 400 ms delay
This study uses its compact SEEG pair features, sound targets, fixed passage split, raw range records, retrieval checks, and source notices. It tests one shared time filter on the same later passages.
usesTraining for one known feature group loss increased mean worst-case skill and decreased mean clean skill in nine SEEG sound readers
This study uses the compact SEEG power arrays, sound targets, passage lists, and source identities. It fits eight raw pair readers and tests exact loss training. The 100 later passage records are reused.
usesA ridge-leverage gate showed no clear mean skill gain for later sound estimates from nine SEEG recordings
Used its compact bipolar power features and source extraction methods through the immediate input package. Fresh retrieval checked all 154 pinned neural ranges.
usesSix calibration passages gave a mean 6.5% lower later squared error than two in nine SEEG sound readers at a 400 ms delay
We used the compact neural features, audio targets, source records, and MIT loading code. We fitted new readers and checked inherited inputs against its manifest.
usesA shared time filter changed mean SEEG sound amplitude skill by −0.00018 in nine listeners at a 400 ms delay
This study uses its compact SEEG pair features, sound targets, fixed passage split, raw range records, retrieval checks, and source notices. It tests one shared time filter on the same later passages.