Participant-Level Results and Trial-Level Predictions for Cross-Dataset Motor Imagery EEG Decoding Using Euclidean Alignment

This dataset accompanies the research manuscript entitled “Unlabeled target EEG improves cross-dataset motor imagery decoding through Euclidean alignment”, submitted to Scientific Reports. The study investigates the effects of Euclidean alignment (EA) on cross-dataset motor imagery electroencephalography (EEG) decoding, with particular emphasis on label-free target calibration, matched evaluation protocols, and participant-level variability. The primary analysis covers four publicly available motor imagery EEG datasets—BCI Competition IV Dataset 2a, Lee2019 (OpenBMI), Cho2017, and PhysioNet EEGMMIDB—comprising 222 participants and 28,184 analysed trials. The deposited data include: Participant-level classification results and paired performance comparisons from the four-dataset reference analysis. Results for zero-target transfer, full-target Euclidean alignment, and limited unlabeled calibration using 5, 10, or 20 trials. Participant-level results from a separate, post-hoc comparison of a log-Euclidean minimum-distance-to-mean (MDM) classifier and a compact EEGNet-style neural network. Trial-level model predictions, calibration and evaluation split memberships, and associated performance summaries for the exploratory classifier benchmark. Supporting documentation and file-integrity information to facilitate inspection and numerical reanalysis. The exploratory MDM/EEGNet benchmark was evaluated on 109 PhysioNet participants, using models trained exclusively on the three development datasets. Its findings should not be interpreted as independent prospective validation. A subsequent raw-data replay reproduced the benchmark using the same code, random seeds, and software environment. Data access and reuse: This repository contains derived research results and prediction data, not the original raw EEG recordings. The underlying EEG datasets remain available from their respective original providers, as cited in the accompanying manuscript. The deposited materials support transparency, verification of the reported analyses, and further investigation of participant-dependent effects in cross-dataset motor imagery EEG classification.