A SEEG power classifier trained on spoken trials gave 6.83 percent balanced accuracy for 20 imagined word items in one participant
We tested whether a fixed SEEG power classifier trained on spoken words could identify the same word items in imagined trials. Public VocalMind data supplied one participant, 20 word items, 110 contacts, and 357 trials in spoken, mimed, and imagined modes. Each test repetition number was absent from training. Spoken to imagined balanced accuracy was 6.83 percent on 119 trials. The chance mean was 5 percent. A test with 9,999 word-mapping permutations gave P = 0.1806. The descriptive 95 percent interval from word resampling was 2.67 to 11.33 percent. Imagined within-mode balanced accuracy was 5.00 percent, with a label-shuffle P value of 0.564. This prevents an inference that transfer alone caused the low score. Spoken and mimed within-mode scores were 17.50 and 18.50 percent, each with P = 0.002 in supporting label controls. Removal of mean spectral level did not establish imagined word decoding. The package preserves all trial features, 36 fitted models, all class scores, source hashes, a raw-data retrieval recipe, source code, numerical checks, and four figures. The source conversion sorts trial files by name, so this study does not treat file order as recording time. Shared visual cues and one participant limit the interpretation. The result applies to these fixed features and this classifier. It does not establish useful imagined speech decoding, a live interface, or a stimulation effect.
stereotactic EEG · imagined speech · speech decoding · cross-mode transfer · ridge regression · balanced accuracy
References
3- usesdoi:10.82901/nemar.nm000251
The analyzed BrainVision source files came from the NEMAR VocalMind conversion. Their exact URLs and SHA-256 hashes are recorded in original/provenance/bids_raw_files.json; the saved dataset description lists version 1.0.0.
- usesdoi:10.5281/zenodo.14696348
This is the source VocalMind release from which the NEMAR BIDS files were converted. The publication records the Zenodo release and the BIDS source URLs and hashes; the DOI is cited as the underlying dataset record.
- doi:10.1038/s41597-025-04741-2
Cited by
6 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
Its VocalMind retrieval and SEEG log power feature method is the source method for the input feature package. We preserve its source record and license. We use its data and feature method, not its cross-task decoding result.
- usesEarly spoken SEEG gave 13.2% mimed word accuracy in one participant without mimed training trials
Used its original public VocalMind feature method and full-trial transfer benchmark through the immediate input publication. The package preserves the source credits and license.
- usesContact mean removal retained spoken and mimed word label association for a fixed classifier in one SEEG participant under exchangeable labels
This study uses its feature data, trial metadata, source readers, and retrieval records. All features were independently rebuilt from complete pinned public raw files.
- usesAdded speech trials did not show a clear gain for a fixed imagined word decoder in one SEEG participant
We used its verified VocalMind features, trial metadata, and extraction code. These files enabled the new calibration test.
- usesSpoken and imagined task SEEG patterns matched for 19 words in one participant under exchangeable word labels
We used its VocalMind feature array, trial metadata, raw retrieval manifest, and source readers. We rebuilt all trial features from the pinned public NEMAR raw files. The new analysis tests word pattern products across different repetition numbers. It measures alignment rather than classifier accuracy.
- usesSix SEEG time windows gave higher spoken and mimed word accuracy than their average in one participant under repetition splits
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.