For fixed sounds, acoustic ranks gave mean gains but failed a shared majority test in 20 iEEG patients and 24 fMRI adults

We tested whether acoustic ranks improve predictions for a sound class omitted from fitting. Separate public studies supplied signed mean local field potential from 20 reserved iEEG patients and BOLD fMRI from 16 reserved adults. Eight adults from a second fMRI study supplied a required study check. Development used five other patients and four other adults. Each person had separate models. Source-labelled music was excluded. Training-only transforms, nested class exclusion and spatial mean removal limited several confounds. Mean rank-minus-raw gains were 1.741, 1.462 and 1.170 percentage points of the class-centred response score. Their 95% person intervals were [0.571,2.912], [0.869,2.055] and [0.362,1.979]. Gains occurred in 12/20, 15/16 and 7/8 people. The fixed shared mean-and-majority rule failed, p=0.251722. The electrical person-and-sound resample range crossed zero. Thus no shared advantage for a typical person or a new sound population was established. A clipping control and known rank invariance separate extrapolation limits from accuracy. A fold-specific tail stress test kept training values and sound order fixed while changing raw predictions. The contribution is a reproducible non-music class-exclusion benchmark, limited positive mean estimates and an explicit failed stronger test. Clinical sampling, different descriptors and class labels, source processing and MRI row-order assumptions limit scope. No common neural mechanism or cause is established. The package gives code, permitted electrical inputs, source versions and terms, local plans, all results, figures and repeat instructions.

human neuroscience · iEEG · fMRI · auditory cortex · acoustic features · rank transform · prediction · distribution shift · negative result · reproducibility