Sound band differences gave no clear mean R² gain in nine depth recordings with a common average reference and fixed ridge training

We tested whether sound band differences improve later high-gamma prediction beyond mean log-mel sound level. We used 325 depth contacts from nine people in OpenNeuro ds005574, version 1.0.2. An early level model selected 68 contacts. Fixed ridge models used the first five minutes of one story. We tested predictions in its last ten minutes. The added features changed mean R² by -0.00343. The 95% participant bootstrap interval was -0.00992 to 0.00340. The exact two-sided sign-flip p-value was 0.3828. Three of nine people had a positive gain. A higher penalty gave a small positive mean gain. The result does not show that frequency information is absent. It applies to this training rule and these recordings. The package preserves compact data, source hashes, code, fitted models, predictions, and repeat checks. It reuses a known cohort and provides a controlled comparison method for later acoustic baselines.

high-gamma amplitude · intracranial recordings · acoustic features · log-mel spectrogram · ridge regression · temporal prediction · depth contacts · model comparison