解码方法
计算机科学
人工智能
计算机视觉
运动表象
神经解码
冲程(发动机)
传感器融合
融合
模式识别(心理学)
图像融合
编码(内存)
神经生理学
语音识别
人工神经网络
图像(数学)
保险丝(电气)
作者
Yunjeong Jang,Yun Kwan Kim,Wanjoo Park,Yun‐Hee Kim,Minji Lee
标识
DOI:10.1109/tim.2026.3706166
摘要
Reliable motor imagery (MI) decoding from electroencephalography (EEG) remains challenging due to intersubject variability, limited training data, and uncertainty inherent in noninvasive neural signals. While recent machine learning approaches have improved accuracy, less attention has been given to decision confidence and robustness in neural interfaces. This study presents a hierarchical connectivity-based MI decoding framework designed to support more reliable neural decoding under data-constrained conditions. We extract phase-based functional connectivity (FC) features and select discriminative connections through a two-stage nonparametric statistical procedure. Dempster–Shafer evidence theory is then employed to model and fuse decision evidence while explicitly capturing uncertainty. In addition, few-shot calibration enables subject-specific adaptation when only limited labeled trials are available. Experiments on a four-class MI-EEG dataset suggest that the proposed method may reduce decision uncertainty and improve prediction consistency across repeated evaluations. In addition, the proposed framework achieved competitive classification performance compared with conventional classifiers under limited-data conditions. These findings suggest that principled uncertainty modeling and hierarchical neurophysiological feature selection may provide measurable benefits for more reliable MI decoding beyond accuracy alone. Overall, this work contributes a structured approach for uncertainty-aware EEG-based decision-making and underscores the importance of confidence quantification in advancing more dependable neural-signal measurement systems. The proposed framework represents a step toward more dependable and interpretable brain–computer interfaces (BCIs) for data-constrained clinical settings.
科研通智能强力驱动
Strongly Powered by AbleSci AI