Word surprisal reduced test error in 3 of 9 patients for a fixed model of depth signals during one story

We tested public depth-contact data from 9 patients in OpenNeuro ds005574 version 1.0.2. All patients heard one story. We used 317 good depth contacts and selected ten per patient with a separate validation segment. The fixed ridge model used acoustic and timing controls. The response was log high-gamma amplitude from 0.2 to 0.6 seconds after each word. After supplied GPT-2 XL word surprisal was added, the mean reduction in normalized test error was -0.000194. Positive values mean improvement. The patient bootstrap interval was [-0.000678, 0.000270]. The exact sign-flip p value was 0.4648. 3 of 9 patients improved. The baseline mean test R-squared was -0.185. Poor baseline performance limits interpretation. These data do not show that the brain lacks a surprisal response. The package preserves compact depth responses, predictions, source hashes, code, and a full retrieval recipe.

word surprisal · high-gamma activity · depth electrodes · natural language · ridge regression · neural encoding