Event and excerpt validation gave 48.9% and 49.2% mean sound class accuracy for an eight-pair SEEG decoder in ten people

We compared event folds with whole-excerpt folds for a fixed SEEG sound class decoder. The public source was OpenNeuro ds004703, version 1.1.0. Both rules tested the same 3,516 selected vowel and fricative events from ten people. Each person had 20 five-fold schedules. Pair selection, scaling, and fitting used training data only. Mean balanced accuracy was 48.94% with event folds and 49.23% with excerpt folds. The paired difference was -0.30 percentage points. Its 95% person bootstrap interval was -1.39 to 0.50 percentage points. The exact two-sided sign-flip p value was 0.7285. This result does not establish equal performance. All-pair models gave about 97% mean training accuracy and 50% mean test accuracy under the fixed penalty. The package adds a common-event validation test, three figures, inputs, predictions, 7,000 fitted models, source hashes, and reproduction code. Independent solver and fresh signal range checks passed. The result applies to this reused listening sample and fixed model. It does not test intended speech, online control, or brain stimulation.

stereoelectroencephalography · brain-computer interfaces · speech decoding · cross-validation · sound classification · ridge regression