In exploratory tests, acoustic feature clipping lowered baseline error in 20 iEEG patients and 24 fMRI adults; a fixed slope test failed

We tested whether four loudness slope summaries improve response predictions for omitted sounds beyond 84 other acoustic descriptors and source classes. Separate studies supplied signed iEEG voltage and BOLD fMRI. A local lock preceded tests in 20 patients, 16 B2021 adults and eight NH2015 adults. Fixed residual ridge gains were 0.0050%, 0.1716% and 0.2518% of target energy. The shared mean-and-sign rule failed, p=0.494083. Sound-weight refit ranges included zero. Every iEEG baseline predicted worse than a constant. A later exploratory repair clipped acoustic features at three training standard deviations. Baseline error then fell in all 44 test people. Mean normalized risks changed from 4.436 to 0.919, 0.879 to 0.610 and 0.907 to 0.663. Pointwise person intervals for the decrease were positive in each study. The clipping shared p was 0.00390625; its Holm value in a conservative final 13-claim family was 0.05078125. The strict shared rule still failed. This model repair has no independent endpoint confirmation. Two sounds supplied 75.5% of original iEEG baseline error. No sound was removed. Independent solvers checked all 53 primary people. Source reconstruction matched every retained voltage mean. The package preserves code, permitted inputs, source pins, plans, results and figures. Clinical sampling, source selection and different signals limit neural interpretation. No common neural slope code or cause is established.

human neuroscience · iEEG · fMRI · natural sounds · loudness slopes · prediction · negative result · acoustic feature clipping · replication · reproducibility