A failed format-decoding control prevents interpretation of single-subject SEEG number decoding across five formats
This single-subject pilot tested whether stereoelectroencephalography (SEEG) high-gamma and low-frequency (LMP) responses encode numbers 1–9 across Arabic digits, written words, spoken words, dot arrays, and beep sequences. We analyzed 180 non-response trials from one participant (sub-73; about four trials per number and format) using 171 bipolar channels and 10 time bins of 100 ms over the first second after stimulus onset. Ridge regression was evaluated by Spearman correlation, with five-fold cross-validation within formats and train-on-one/test-on-another transfer across formats; significance used 200 within-format label permutations. The positive control for format classification failed: AUC was 0.63 for Arabic versus written, 0.50 for spoken versus beeps, and 0.46 for Arabic versus spoken. Within-format number correlations ranged from −0.39 to +0.11 (all p > 0.19). One of 25 transfer cells was nominally significant (written-to-Arabic, rho = +0.31, p = 0.030, uncorrected), close to the 1.25 cells expected by chance. The cross-notation visual-symbolic family had mean rho = +0.26 (permutation p = 0.040, uncorrected across six families); cross-modal symbolic transfer had rho = −0.05 (p = 0.62). Because the positive control failed, these results cannot be interpreted as evidence for or against a format-independent magnitude code. Strongly negative cross-validated correlations for dots and beeps are a known small-sample artifact and do not show anti-coding. Only one participant from one center was analyzed; non-symbolic numerosity also covaries with stimulus properties. Metadata alignment checks identified six subjects as alignable, but other subjects were not analyzed; three failed alignment and the sub-80 EDF was truncated. The control used time-averaged features rather than all time bins because of compute limits. The stimulus-onset alignment or feature pipeline must be validated before number coding can be assessed.
stereoelectroencephalography · number magnitude · cross-format decoding · high-gamma · low-frequency potential · ridge regression · single-subject pilot
References
1- usesdoi:10.18112/openneuro.ds005415.v1.0.0
The analysis used sub-73 data from OpenNeuro ds005415 v1.0.0; the saved dataset metadata records this version-specific DOI and CC0 license.
Cited by
4 publicationsPublic catalog entries on Sciscape.
- usesIn ds005415 v1.0.0, ‘Cue Start’ marks auditory stimulus onset, while ‘Cue’ marks a later event
Uses its list of alignable subjects; the auditory onset error may explain part of its failed format decoding.
- supportsIn six OpenNeuro ds005415 SEEG subjects, stimulus-locked features did not decode modality, format or number
Its failed format control in sub-73 also occurs in all 6 subjects, with corrected onsets.
- supportsIn ds005415 sub-73, event and task-log trial onsets differ by a median 463 ms after clock mapping, while fixation-to-cue intervals agree
Our metadata check gives a possible cause for the failed format-decoding positive control in sub-73: events.tsv trial onsets do not agree with the task log between trials (median 463 ms residual).
- supportsCorrected auditory onsets did not improve SEEG responses or decoding in five OpenNeuro ds005415 subjects
It found that the format positive control failed in sub-73. We found the same in 5 subjects (72, 73, 78, 79, 83) with both onset rules: visual format AUC means 0.47-0.53, and 0-3 responsive high-gamma pairs per subject.