Six SEEG time windows gave higher spoken and mimed word accuracy than their average in one participant under repetition splits
We tested whether a word reader could use the average of six SEEG time windows. We used public VocalMind data from one participant, with 19 words, six repetitions, and 342 trials. We compared two ridge readers under nested repetition splits. The reader with separate windows used 1,980 inputs. The average reader used 330 inputs. Accuracy was 17.5% versus 10.5% for spoken words and 20.2% versus 8.8% for mimed words. The separate readers passed shuffled-label tests in both tasks after correction for six tests. Their higher accuracy also occurred at all six tested fixed penalties. Imagined accuracy was 10.5% versus 6.1%. That difference depended on one word. These results apply to the specified readers and trials. The test cannot separate speech activity from visual cue activity. It does not establish performance across participants or brain stimulation effects. The package contains three figures, input features, fitted readers, predictions, code, provenance, and public source retrieval instructions. Full analysis results matched exactly in a fresh package copy.
stereoelectroencephalography · speech decoding · brain-computer interface · ridge regression · time windows · VocalMind · word classification
References
5- usesSpoken and imagined task SEEG patterns matched for 19 words in one participant under exchangeable word labels sci:sha256:de42a84ec68163186569927aebfbc2d96c47d125992ca12d5322608d3dc44a3d
We used its public VocalMind feature arrays, trial labels, raw audit arrays, pinned source metadata, and full raw validation record. The new analysis compares separate time windows with their average under nested repetition splits. We checked three public raw trials again in this run.
- usesA SEEG power classifier trained on spoken trials gave 6.83 percent balanced accuracy for 20 imagined word items in one participant sci:sha256:acba984d9e9a4769346f785ddea358590e3b31967715ff505e64b14d22cbe78b
We used its log power feature definition and ridge word reader design as source methods. We added nested penalty selection and label tests for both feature forms. The source method and license notices are preserved.
- doi:10.1038/s41597-025-04741-2
- usesdoi:10.5281/zenodo.14696348
Source VocalMind data record; the publication identifies Zenodo 14696348, version 2, as the public source record.
- usesdoi:10.82901/nemar.nm000251
The analyzed BIDS conversion is NEMAR nm000251, version 1.0.0, at the pinned commit recorded in the source metadata.
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesA rank three SEEG reader used 6016 coefficients and gave 11% to 12% word accuracy under repetition splits in one participant
We used its feature arrays, trial labels, pinned raw retrieval code and metadata, repetition splits, and saved full ridge and time-average predictions. These inputs let us test a fixed rank three reader on the same trials. All 684 comparison predictions matched. The new work adds reduced rank fits, a storage audit, a streaming reader, and full raw reconstruction checks.
- usesEarly spoken SEEG gave 13.2% mimed word accuracy in one participant without mimed training trials
Used its VocalMind feature file, trial table, raw retrieval and reconstruction code, source metadata, and raw audit trials. The present study adds source-only model selection across separate source and target time windows.