At 16 sites, template noise subtraction reduced sound identity accuracy in six SEEG patients and 24 fMRI adults

We tested whether subtracting estimated template noise improves neural sound identity matching. Separate public studies supplied SEEG high-gamma responses and BOLD fMRI estimates. Development used two patients and four MRI adults. A local code lock preceded tests in six other patients, 16 B2021 adults and eight NH2015 adults. Each reader used ten 16-site panels. Training sounds set site scales. Two repeats supplied templates; the other repeat supplied queries. Full template noise subtraction reduced accuracy in all 30 reserved people. Mean changes were -5.791, -3.176 and -3.259 percentage points. Their 95% person Student intervals were [-7.000,-4.582], [-3.782,-2.570] and [-3.862,-2.656]. The fixed three-cohort intersection rule passed, p=0.015625. False choices favored templates with greater repeat disagreement. Fixed site, spatial and covariance checks kept negative mean changes. Short SEEG baseline subtraction greatly lowered both accuracies and reduced the loss to 0.667 points. Two B2021 adults improved with one-quarter correction. For one noiseless query, an exact independent-error model had correct expected distances but corrected accuracy 2^-K against K false candidates; ordinary accuracy was one. The model was not fitted to people. A second calculation reproduced every primary accuracy. Source audits checked all retained neural values. The package has code, source pins and terms, all outcomes, plans, figures and repeat steps. Original neural arrays remain external. Clinical sampling, source screens, separate cohorts, fixed sounds, repeat dependence and measurement filters limit scope. This is a shared reader-performance result. It does not establish a common neural code, noise source or cause.

human neuroscience · SEEG · fMRI · sound identification · template noise · crossvalidation · bias and variance · reproducibility