EEG-based mental workload (MWL) classifiers consistently achieve high within-dataset accuracy but collapse when applied across datasets recorded under different paradigms or hardware. This cross-domain generalisation gap limits real-world deployment of passive brain-computer interfaces. We evaluate transfer strategies across three publicly available EEG-MWL datasets — CogBCI (29 subjects, 3 sessions), Neuro2021 (15 subjects), and STEW — revealing a mean within-domain accuracy of 78.8% versus cross-domain accuracy of only 44.0%, a gap of 34.8 percentage points. We propose Multi-Source Domain Generalisation with Few-Shot Fine-Tuning (MSDG-FT), which reduces this gap to 6.6 percentage points using as few as 50 labelled calibration samples. Cross-session drift on CogBCI is further characterised across all six session-pair directions, showing near-chance baseline accuracy (36.0%) that recovers to 51.6% with minimal calibration. • A 3 × 3 cross-domain transfer matrix quantifies generalisation failure across three heterogeneous EEG-MWL datasets and establishes a reproducible benchmark for future methods. • Multi-source pre-training combined with few-shot target-domain fine-tuning (MSDG-FT) closes the 34.8% transfer gap to 6.6% using only 50 labelled samples from the target domain. • Random calibration (20 samples) matches sophisticated confidence-weighted selection (p=0.28), demonstrating simple baselines suffice. Cross-session benefits vary by dataset: CogBCI +15.6%, Neuro2021 +3.5%, indicating task-dependent effectiveness.
G et al. (Wed,) studied this question.