Sound labels lowered test error for averaged responses in 20 iEEG patients and 16 fMRI adults; separate BOLD fits gave no clear gain

We tested whether source sound labels improve prediction of responses to sounds kept out of fitting. Separate public human studies supplied signed iEEG repeat means and BOLD fMRI. Four people per study supplied development. Fixed tests used 20 other patients and 16 other adults. The baseline combined linear and nonlinear acoustic kernels. A category kernel and an acoustic clustering control used fixed weights and equal group counts. Mean category gains were 2.434 [0.886, 3.981] and 3.342 [2.168, 4.516] percentage points against intercept error. Brackets give 95% person Student limits. Advantages over acoustic clusters were 2.418 [0.553, 4.283] and 7.577 [6.590, 8.564] points. Both shared rules passed Holm correction, p=0.011818 and 0.020695. However, all three separately fitted BOLD estimates had gain intervals across zero, although all beat acoustic clusters. An exact fixed-prediction identity and a later audit separate test averaging from refitting. A further eight-person fMRI acquisition check had mean gain 1.311 points; its common-test-sound range crossed zero. An acoustic-only construction gave both gains without a separate category information channel. Thus, the finding concerns these processed mean responses and prediction rules. It does not establish information independent of all acoustics or a common neural cause. Cohort, label, sampling, filter and source-selection differences limit scope. The package gives licensed electrical means, source pins and terms, code, all results, figures, audits and repeat instructions. Original MRI arrays remain external.

human auditory cortex · intracranial EEG · fMRI · sound categories · acoustic models · kernel regression · repeat averaging · prediction error