Minima Possible Weights: A Homogenous Deep Ensemble Method for Cross-Subject Motor Imagery Classification

• Published in: IEEE access Vol. 13; pp. 29134 - 29146
• Main Authors: Dinh, Quang Pham Lam, Nambu, Isao
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Adaptation models; Brain modeling; Brain-computer interface; Calibration; Classification; cross-subject; Data models; Datasets; deep ensemble learning; Deep learning; Ensemble learning; Human-computer interface; Imagery; motor imagery; Motors; multi-class classification; Neurons; Predictions; Predictive models; test-time adaption; Testing time; Training; Unsupervised learning
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2025.3540164

Motor Imagery (MI) systems in Brain-Computer Interface (BCI) research provide communication and control solutions for individuals with motor impairments, yet cross-subject classification remains challenging due to substantial inter-subject variability. In this study, we propose the Minima Possible Weights (MPW) method, an unsupervised learning approach designed to enhance MI classification through ensemble deep learning. MPW aggregates predicted probabilities from multiple models and selects the class with the lowest associated weight for the final prediction. We evaluated MPW against various ensemble learning and test-time adaptation methods using two benchmark datasets: BCI Competition IV Dataset 2a and PhysionetMI. Our results indicate that MPW achieves cross-subject classification accuracy of up to 64.75% on BCI Competition IV Dataset 2a and 66.92% on PhysionetMI. Although the current performance is not yet sufficient for practical BCI applications, MPW shows potential in reducing calibration time and easing the burden of adapting models to new subjects.