Centered late high gamma map shrinkage under a common average reference gave no lower later mean map error than zero maps in nine SEEG patients

We tested whether shrinkage of high gamma contact maps reduces prediction error on later speech passages. We used public OpenNeuro ds004703, version 1.1.0. Nine patients supplied 286 English passage occurrences and 1178 good SEEG contacts. We reused archived features from patients with low map transfer in an earlier test. The first half of passages fitted each map. The later half supplied the test mean map. The primary measure was centered high gamma change at nominal 10–12 seconds, with a common average reference. The fixed estimate reduced mean patient error by 51.5 percent relative to raw means. The 95 percent patient bootstrap interval was 39.6–62.6 percent. The exact two-sided sign-flip p value was 0.00390625, with Holm adjusted p value 0.0078125. These tests require independent patient effects and a symmetric null. No patient had lower error than a zero map. Six estimates were zero maps. Three had higher error. Mean patient relative error was 4.9 percent higher than zero. The raw-mean improvement therefore did not show added contact information in this test. It does not show an absence of speech responses. The package preserves the plan, code, features, fitted maps, errors, source versions, range hashes, figures, and reproduction instructions. Fresh pinned raw-data checks matched all 1144 windows.

stereo-electroencephalography · high gamma · speech · map shrinkage · prediction error · patient-level analysis