Subtracting predicted sound responses lowered mean coupling in exploratory tests of nine iEEG patients and 25 fMRI listeners

We tested human neural responses to two different English stories in separate iEEG and fMRI studies. A fixed 100 ms sound spectrum-change feature failed the planned shared prediction test. Later exploratory tests found that subtracting responses predicted by the fixed baseline sound model lowered measured coupling. Models fit the first 60% of each story and predicted the last 40%, with guard gaps. The mean Fisher-z decrease was 0.03565 in nine iEEG patients. Its 95% person bootstrap interval was 0.02078 to 0.05121. The mean decrease was 0.23756 in 25 fMRI listeners. Its interval was 0.18932 to 0.28888. Eight of nine patients and all 25 listeners had a decrease. The shared sign-flip p was 0.003906; Holm correction across five logged shared questions gave 0.01953. This assumes symmetric independent person effects at the null. The corresponding corrected shared binomial sign p was 0.09766. Test halves and an iEEG distance check kept positive mean changes. Joint time shifts gave diagnostic fractions of 0.002 in each measure. There is no independent person confirmation. Contact selection, fixed stories, clinical sampling and common reference limit scope. Spatial and temporal samples differ. Cleaning uses the full run. The result shows that a sound-aligned model component can change measured coupling. It does not identify direct communication, recover true connectivity, or establish causality. The package preserves all failed and discordant tests, code, compact inputs, source versions and terms, audits, figures and repeat steps.

human neuroscience · iEEG · fMRI · natural speech · functional coupling · common input · prediction · exploratory analysis · negative results