Vocal labels failed shared forecast tests in 14 iEEG patients and 16 fMRI adults

We tested whether a human-vocal versus other-sound label can replace five classes in a short response forecast. Separate public studies supplied signed human iEEG means and BOLD fMRI. Four people per primary study supplied development. A local code lock preceded new endpoints in 14 patients and 16 MRI adults. Mean binary gains over acoustic neighbors were 1.340 points [-0.139,2.819] and 3.220 [2.687,3.752]. Brackets give 95% person Student limits. One point is 1% of constant-forecast squared error. The fixed shared usefulness rule failed, Holm p=0.072126. The fixed half-benefit rule failed, Holm p=1. In all 16 primary MRI adults, binary gain was less than half the five-class gain. An added eight-adult MRI study gave gain 1.909 [0.122,3.696] and failed the half-benefit rule. Regional differences were positive in both primary studies. A later nine-test audit gave shared regional p=0.013840; it is exploratory. Exact subset risk exposed the benefit of pooling 16 rather than four nonvocal examples. One-site gain intervals crossed zero in both modalities. Pre-event electrical gain was positive but absolute skill was negative. Without shrinkage, electrical gain rose to 13.927 points while absolute skill was uncertain. A separate matrix solver checked all five risks in all 38 test people. The package supplies the assay, permitted compact inputs, all negative results, source checks and terms, figures, and repeat steps. It does not confirm a shared positive forecast under the fixed rules, one sufficient neural dimension, or a causal effect. Clinical sampling, distinct cohorts and libraries, source screens, measurement filters and limited electrical precision restrict scope.

human neuroscience · intracranial EEG · fMRI · auditory cortex · voice · prediction risk · label compression · negative result