Random splits gave higher mean speech and music block accuracy with eight contact pairs in six SEEG recordings of one film
We tested a sparse reader of speech and music block labels in six public SEEG film recordings. The reader used eight bipolar contact pairs and half-second power features. Mean balanced accuracy was 65.17% for random window splits and 61.59% when a full 60-second section pair was held out. The observed difference was +3.58 percentage points. A participant bootstrap gave a 95% interval from 2.30 to 5.32 points. Both main tests used the same train and test counts. Selection and scaling used training data only. A separate later-half test gave 60.63% mean accuracy, with less training data. The known event schedule alone gave 100% block accuracy. Thus these scores do not prove a signal specific to speech. The package contains checked source hashes, derived data, all test predictions, fitted models, figures, and code. The result applies to this reader, six recordings, and one film. It does not measure speech intent or brain stimulation.
SEEG · speech and music blocks · brain-computer interfaces · validation · random window splits · held-out sections · balanced accuracy · electrocorticography
References
2- usesdoi:10.18112/openneuro.ds003688.v1.0.7
Version 1.0.7 supplied the public SEEG signals and block events used to build this validation comparison.
- supportsdoi:10.1038/s41597-022-01173-0
The source data paper documents the dataset and its audiovisual film stimulus, which define the scope described here.
Cited by
3 publicationsPublic catalog entries on Sciscape.
- Full section tests did not detect a mean block accuracy change after 96–104 Hz bin removal in six SEEG recordings of one film
[Legacy relation: reproduces] Independent full-signal retrieval and extraction reproduced all four original feature arrays and all 18 original participant accuracy scores across section, forward, and random splits.
- usesFull section tests did not detect a mean block accuracy change after 96–104 Hz bin removal in six SEEG recordings of one film
Uses the six-person cohort, source metadata, extraction and sparse-reader code, original features for exact comparison, and original scores. New spectra and frequency exclusions were extracted again from complete public signals.
- usesMixed windows gave 9.1 percentage points higher mean balanced accuracy than separate records for a four-pair seizure reader in 28 SEEG subjects
Its matched-count comparison of mixed windows and separate film sections informed the validation design. This study tests separate seizure records with a different fixed reader.
- Adding eight SEEG pairs changed mean later probability error in opposite directions under four fixed penalties in six recordings of one film
[Legacy relation: reproduces] A new retrieval of all six complete raw signal files exactly reproduced all 1,053,312 inherited SEEG feature values. This check applies to the feature data, not to every analysis result in the cited publication.
- usesAdding eight SEEG pairs changed mean later probability error in opposite directions under four fixed penalties in six recordings of one film
This study uses the fixed six-person cohort, SEEG features, source metadata, and retrieval code from this publication. It adds source EOG and ECG inputs and a later-time probability test.