A Harmonized Multi-Dataset EEG Benchmark for Cross-Hardware Generalization

This repository contains the harmonized cross-device EEG dataset generated to evaluate the resilience of deep learning architectures against real-world inter-device distribution shifts. The benchmark aggregates, temporally segments, and spatially harmonizes five distinct public motor imagery (MI) EEG corpora into a standardized 20-channel format:1. PhysioNet Motor Imagery2. Cho20173. Lee2019-MI4. Shin2017A5. Weibo2014 Dataset Details: - Task: Binary Motor Imagery (e.g., Left Hand vs. Right Hand execution/imagination).- Channels (20): Fp1, Fp2, F7, F3, Fz, F4, F8, T7, C3, Cz, C4, T8, P7, P3, Pz, P4, P8, O1, Oz, O2 (standardized via international 10-20 system and spherical spline interpolation).- Preprocessing: Band-pass filtered (8–32 Hz), resampled to 100 Hz, and segmented into uniform 3-second temporal windows (300 time steps per trial).- Alignment: Mathematically aligned using Euclidean Alignment (EA) to center spatial covariance structures and mitigate baseline cross-hardware variability. The data is saved as a compressed NumPy array (`cross_device_eeg_aligned_FINAL.npz`) containing the aligned trials (X), binary labels (y), and macro-domain indicators (domains) suitable for strict Leave-One-Dataset-Out (LODO) cross-validation.