Leakage-Free Multi-Model Ensemble Learning for Session-to-Session Motor Imagery EEG Decoding

Mitigating performance degradation across recording sessions is essential for brain-computer interface (BCI) technology to move beyond controlled laboratory environments and toward real-world applications. Electroencephalogram (EEG) signals exhibit substantial inter-session variability due to differences in electrode placement, changes in user state, and intrinsic neurophysiological fluctuations, leading to distribution shifts and frequent recalibration requirements. To improve robustness under temporal non-stationarity, we propose a leakage-free multi-model ensemble learning framework that integrates heterogeneous modeling paradigms with complementary inductive biases. A diverse candidate pool is constructed using domain adaptation methods, deep learning-based motor imagery classifiers, and classical spatial-filtering approaches. Ensemble members are selected solely based on source-session validation performance, without access to target-session labels, and their class-posterior probabilities are combined via soft voting. Experiments on a multi-session motor imagery EEG dataset from the Korea Brain Imaging Research Center, using a subject-specific session-to-session transfer protocol, demonstrate the effectiveness of the proposed approach. The proposed method achieves an accuracy of 63.66% ± 14.85% and a macro-F1 score of 62.52% ± 15.81%, while also improving threshold-free performance measures on average compared with individual models. These results indicate enhanced performance and robustness under inter-session variability.