EEG Feature Robustness Under Cross-Dataset Distribution Shift Using Classical Models

This repository presents a machine learning study on EEG (electroencephalography) brain signal classification for motor imagery tasks. The project investigates how well classical EEG feature representations (time-domain, band power, and FFT-based features) generalise across datasets under a subject-aware evaluation protocol. The analysis focuses on realistic deployment conditions by avoiding subject leakage and evaluating zero-shot cross-dataset transfer between the PhysioNet EEG Motor Imagery dataset and BCI Competition IV Dataset 2a. Version 3 introduces major methodological corrections compared to earlier versions, including: - Subject-level separation using grouped cross-validation - Strengthened time-domain feature representations - Feature dimensionality control using PCA - Statistical evaluation using permutation tests - Cross-dataset evaluation with bootstrap-based uncertainty estimation Results show that while classical features perform reasonably within datasets, they fail to generalise reliably across datasets. The previously observed ranking of FFT > band power > time-domain features is not supported under the corrected evaluation protocol. This work serves as a reproducible baseline highlighting the challenges of cross-dataset generalisation in EEG-based machine learning systems.