Linear feature replacement increased mean sound error after simulated group loss in ten SEEG listeners with eight pairs and 400 ms delay
Can early SEEG correlations replace missing inputs without increasing later sound readout error? We tested OpenNeuro ds004703 v1.1.0 recordings from ten listeners. Each reader used eight adjacent contact pairs, 70–110 Hz power, and a specified 400 ms output delay. We removed one selected name group at a time and supplied its identity. Fits and choices used early passages; later passages supplied the test. Linear replacement kept the full reader fixed. Mean error skill was 2.33 points with replacement and 5.30 with mean filling, a mean change of −2.97 points. Five people gained and five lost. The person bootstrap interval was −6.11 to −0.32 points; the person and shared-sound interval was −8.71 to +0.94 points and includes zero. Missing-feature error fell in nine people. Removing clipped excerpts gave a −3.22 point mean change. At a matched penalty, replacement equaled a reader fitted to surviving inputs within numerical precision. This exploratory test uses reused same-session excerpts and known simulated loss. It does not establish performance under physical failure, in a live interface, or for brain stimulation. The SCI package retains the report, figures, models, results, and MIT code; source recordings, metadata, and derived input caches are omitted because source terms conflict. The source dataset’s CC0 metadata and README restrictions are both documented; commercial rights are not asserted.
SEEG · sound amplitude · feature replacement · ridge regression · simulated input loss · brain-computer interfaces
References
6- usesJoint SEEG pair selection reduced input correlation but did not establish lower speech error at eight pairs and 400 ms delay in ten listeners sci:sha256:9770658d97c5976c9fcf5dd66bfdc6a68e30b0ed33c2d78776e648c088da9233
Uses its 95 corrected feature excerpts, 39 sound targets, chronological split, source provenance, and range extraction code. We check copied hashes and repeat source extraction. New readers test known group loss.
- usesTraining for one known feature group loss increased mean worst-case skill and decreased mean clean skill in nine SEEG sound readers sci:sha256:b4bc3665fe36a489d98947ecc3dda47437984fd090bd9b7123588dca7bd43fbe
Uses its known feature group loss question as context. The present study tests covariance replacement with new eight-pair readers and one power band.
- usesNormal probability averaging reduced mean Brier error after simulated loss of one contact in eight SEEG readers with 5% covariance shrinkage sci:sha256:de8ed753577c67d1ece54786d9f63885e9a1b4591eb835a39e14475a7980a156
Uses its conditional normal treatment of known missing inputs as method context. The present study tests linear point replacement for sound amplitude, with a fixed reader.
- usesRidge weights from training data gave mixed error changes under known input loss in eight SEEG recordings sci:sha256:4ba1ff6fb8895dcb1ae3f57af8bed1fa89a122936ba779be380b9c612d899252
Uses its ridge survivor-weight adjustment as context for the refit control. The present continuous-output benchmark verifies equivalence to covariance replacement at alpha=lambda/n.
- usesdoi:10.18112/openneuro.ds004703.v1.1.0
The analysis used OpenNeuro ds004703 v1.1.0; the version-specific DOI is recorded in the saved dataset metadata.
- doi:10.1038/s41467-024-44844-9
Cited by
1 publicationPublic catalog entries on Sciscape.
- usesEight SEEG pairs did not show lower estimated pitch error beyond sound level in periodic speech windows from ten listeners at 400 ms delay
We used its 95 corrected SEEG feature arrays, 39 sound-level target arrays, source metadata, pinned retrieval records, and extraction code. We made new fundamental frequency targets and fit new readers. The inherited reader outputs were not used as pitch predictions.