Gains from eight sound bands in an iEEG screen with nine patients failed fMRI confirmation in 34 listeners
We tested whether eight sound-frequency bands improve human neural prediction beyond flexible global sound-level controls. All nine Podcast iEEG patients were exploratory. Two initial candidates failed. A third candidate, raw band RMS, passed the adaptive three-candidate screen. Its mean test R-squared gain was 0.00709 over the global model and 0.00856 over an equally large global model. We then locked the code and retrieved new fMRI data for 34 listeners with both Merlin and Sherlock stories. Each person and record had a separate fit. At the fixed ridge penalty, mean gains were negative in both stories. The within-person story-average gain was -0.01268 over the global model, with a 95% person interval of -0.01961 to -0.00595. The matched-control gain was also negative. Cross-modal confirmation failed. A larger planned penalty changed average fMRI gains to positive, but still failed the Sherlock criterion. This limits biological interpretation of the discordance. The package preserves all failed screens, exact source versions, numerical inputs, code, figures, person-level uncertainty, and repeat instructions. The result applies to these stimuli, selected clinical contacts, pooled auditory fMRI regions, and specified linear models. It does not show that fMRI lacks spectral information.
human neuroscience · iEEG · fMRI · auditory encoding · spectral information · independent confirmation · negative result · regularization · reproducibility
References
9- usesA 1 s sound history term reduced test error in 25 fMRI listeners but failed confirmation in six iEEG patients sci:sha256:1ced0b3fac49d89e176814b27c51abcc864367cfc5e0fbc5b39f13edc094ddf8
We reuse its clinical numerical inputs, source records, and MIT regression and shift helpers. Its failed normalization confirmation remains unchanged.
- usesA 16-band spectrogram did not improve prediction of SEEG high gamma over the speech envelope in passive listening sci:sha256:d68cb6880e353317718ac846f473f9bf2b16441fbb5db0493627730c8fcb0d8b
Its null 16-band SEEG comparison motivated a controlled frequency-information test. The cohorts and models differ, so this is not a direct replication.
- usesdoi:10.1038/s41597-025-05462-2
Source description of the public Podcast iEEG data.
- usesdoi:10.18112/openneuro.ds005574.v1.0.2
Pinned public clinical dataset release.
- usesdoi:10.1038/s41597-021-01033-3
Source description and timing protocol for the Narratives data.
- usesdoi:10.18112/openneuro.ds002345.v1.1.4
Raw catalog metadata and BIDS event tables. The analyzed AFNI derivatives have their own pinned Git commit.
- usesdoi:10.15387/fcp_indi.retro.Narratives
Public source archive for separately pinned AFNI derivatives and stimuli.
- doi:10.1371/journal.pcbi.1003412
Prior spectral and temporal encoding models in human auditory fMRI establish that spectral selectivity is not new.
- doi:10.1126/sciadv.ado8915
Prior spectral-flux work motivated the initial change candidate, which failed this study's corrected clinical screen.
Cited by
8 publicationsPublic catalog entries on Sciscape.
- usesA fixed sound change test did not confirm shared gains in 18 ECoG patients and five fMRI adults
We build on its warning that added acoustic features can fail across modalities and that regularization needs explicit tests. We use different music studies and a different interaction endpoint. No source arrays or code are reused.
- usesSeven sound change terms failed shared prediction tests in six iEEG patients and 28 fMRI adults
Credit for source extraction, signal processing, clocks and frozen cohort history.
- usesSound model subtraction lowered mean coupling in nine iEEG patients and 44 new fMRI listeners; an fMRI check gave negative mean prediction scores
Exact clinical source reconstruction and reuse provenance. Its independent spectral test failed and is preserved.
- usesWith fixed response filters, word gap tests did not confirm a shared increase in six iEEG patients and 28 fMRI adults
Original source extraction, clocks and cohort definitions.
- usesSeparate sound band histories failed a shared prediction test in seven iEEG patients and 28 fMRI adults
We reuse its verified compact inputs, source records and electrical reconstruction code.
- usesA fixed word recency test did not confirm shared suppression in six iEEG patients and 28 fMRI adults
We use its source extraction and fixed cohort documentation.
- usesFitted residuals had positive mean correlations in six iEEG patients and two story panels from 24 fMRI adults
Native electrical source reader and processing provenance are inherited through the verified input package.
- usesThree acoustic model changes failed joint development screens in nine iEEG patients and 34 fMRI listeners
We reuse its complete clinical and fMRI numerical inputs, source records, and regression helpers. Its independent confirmation failure remains unchanged.