A smooth acoustic kernel did not pass a shared gain test in 14 iEEG patients and 8 fMRI adults

We tested a fixed smooth acoustic model in two human neural measurements. A prior failed voice-interaction test motivated the rule. The new test used responses to 250 synthetic sounds in 14 patients with intracranial LFP and to 165 natural sounds in eight fMRI adults. Nine patients had been in the prior natural-task analysis. The new electrical task was held out, not a new patient cohort. The fMRI test used a separate study. No voice label entered either model. Six outer sound folds and three inner folds compared linear and Gaussian acoustic kernels. Both centered kernels had the same training trace. Mean gains against intercept error were +0.152 and +2.350 percentage points. Their 95% person Student intervals were -0.110 to +0.415 and +1.049 to +3.650 points. Gains were positive in 8/14 and 8/8 people. The shared rule failed. Two-claim Holm p was 0.79053. The five electrical patients absent from the prior analysis had a negative mean gain. A fixed control used real acoustic features and exactly linear simulated responses. The smooth kernel still gained 2.330 points in the synthetic-sound feature design, with 20/24 positive records. Thus, estimator gain alone cannot identify nonlinear neural computation. Weak electrical prediction, separate cohorts, different features, source selection and unequal precision limit scope. The package gives the failed test, all checks, a new compact electrical derivative, code, source hashes, figures and repeat instructions. No positive cross-modal neural generalization is claimed.

auditory cortex · intracranial EEG · fMRI · acoustic prediction · Gaussian kernel · linear control · negative result · synthetic sounds