A correction from up to four sound class examples lowered mean model error in 20 iEEG patients and 16 fMRI adults

We tested whether errors from a few sounds can correct acoustic predictions for other sounds of the same class. Separate public studies supplied human iEEG local field potential means and BOLD fMRI response estimates. Development used four people per study. A local code lock preceded tests in 20 other patients and 16 other adults. The base model fit other sound classes. A fixed correction added 0.25 times the mean residual from up to four class examples. Each score is the exact expected error over allowed support subsets. Equal-count, acoustic-matched other-class residuals formed a control. Mean gains over the base model were 5.363 [3.416, 7.310] and 9.295 [8.928, 9.662] points of mean-prediction error. Mean advantages over the control were 6.479 [1.749, 11.210] and 4.146 [3.743, 4.548] points. Brackets give 95% person t limits. Both fixed shared claims passed Holm correction, p=0.0000148423 and 0.00493605. Gains were positive in 19/20 and 16/16 people. One-example iEEG gain remained uncertain. MRI transfer between different session estimates stayed positive. Eight adults from another MRI acquisition gave gain 7.709 [6.632, 8.785]. An added exploratory comparison used the same class examples to refit the full acoustic model. The simple correction had greater mean gain by 1.645 [0.786, 2.503] and 4.913 [4.702, 5.125] points. Negative cases and checks remain. The package gives a small calibration benchmark, code, allowed inputs, source hashes, results and repeat steps. The result applies to fixed sound panels, source classes, selected sites and this model. It does not isolate semantic information, brain learning or a common neural cause. Cohort, sampling, source selection and measurement differences limit scope.

human neuroscience · intracranial EEG · fMRI · auditory cortex · residual calibration · sound classes · prediction · cross-modal testing