Pre-registration: Cue-Interval Leakage and an External-Electrode Confound Control in Inner-Speech EEG Decoding — a Pre-Specified Re-evaluation of OpenNeuro ds003626

Abstract

This deposit contains a pre-registration, not results. Two versions are included. Version 1 was written before the dataset's published data descriptor had been read. Version 2 supersedes it and is the analysis plan; both are deposited so that the revision is visible. The plan specifies a re-evaluation of inner-speech EEG decoding on OpenNeuro ds003626 (Nieto et al., Scientific Data 9:52, 2022). Four questions are fixed in advance: how much four-class accuracy falls when the epoch is restricted to the action interval rather than the full stored 4.5 s trial, which contains a direction-indicating visual cue perfectly correlated with the class label; whether action-interval inner-speech accuracy exceeds an empirical permutation chance level under leave-one-session-out evaluation with all fitted transforms scoped to training folds; whether the class can be decoded from the eight external EOG/EMG electrodes alone; and how inner speech compares with the pronounced-speech and visualised conditions recorded in the same participants. Time windows, features, classifiers, fold definitions, permutation procedure, correction method and exclusion criteria are all fixed here. The author commits to reporting the outcome whichever way it falls. At the time of deposit, no data file from ds003626 had been downloaded, loaded, inspected or plotted. Only the published data descriptor and the dataset's public code repository had been read. Code: to be deposited with the results under an open licence.