Quadratic sound terms did not pass a fixed gain test in six iEEG patients and 25 fMRI listeners

We tested whether products of sound-band amplitudes improve later response prediction beyond linear sound bands, global sound level and word timing. Public data came from separate human iEEG and fMRI studies. Three iEEG patients supplied development. A local plan fixed confirmation in six other patients and 25 fMRI listeners. The extension used 36 quadratic terms after removal of the linear training span. Validation selected each penalty and could select zero correction. Mean test R2 gains were +0.001106 in iEEG and -0.000970 in BOLD. Pointwise 95% person and common-time-block resampling intervals were +0.000389 to +0.002003 and -0.003342 to +0.000288. Positive gains occurred in 5/6 and 7/25 people. The fixed shared rule failed; its joint sign p was 0.992683. All fixed checks, losses and zero corrections are retained. The contribution is a new fixed nonlinear prediction test and a compact repeat tool. It does not establish a shared neural mechanism or absent nonlinearity. Clinical selection, region pooling, separate stories, full-run fMRI processing and coarse features limit scope. Code, numerical inputs, source versions and terms, plans, uncertainty draws, figures and repeat steps are preserved.

human neuroscience · iEEG · fMRI · speech · quadratic model · prediction · negative result