A joint acoustic model did not pass a shared gain test in 20 iEEG patients and 16 fMRI adults

We tested whether a joint acoustic model improves sound response prediction beyond separate smooth cue functions. Two public human studies supplied separate iEEG and fMRI cohorts. Four people per modality supplied development. The fixed test used 20 other patients and 16 other adults. Six outer sound folds separated tests from fitting. Three inner folds selected penalties and joint model settings. The joint model included the additive reference as a candidate. Mean gains were +3.457 and +0.761 points of intercept error. Their 95% person Student intervals were [-2.497 to +9.412] and [+0.637 to +0.886]. Gains were positive in 7/20 and 16/16 people. The fixed shared rule failed; conjunction sign p was 0.942341. A few large electrical gains raised its mean. Artificial additive responses also gave joint-model gains on both acoustic panels. This counterexample limits a neural interaction interpretation. This result tests a broader joint kernel than an earlier fixed cue-pair comparison. The package preserves code, the permitted electrical derivative, source hashes, all errors, controls, figures and repeat steps. These are new endpoint tests in reused cohorts. They are not new-study replication. Different features, sounds, sampling, cohorts and measurement filters limit transfer. No positive cross-modal neural claim or causal effect is established.

human neuroscience · intracranial EEG · fMRI · auditory models · additive kernels · nested validation · negative result