LGNet: Learning local–global EEG representations for cognitive workload classification in simulated flights

• Published in: Biomedical signal processing and control Vol. 92; p. 106046
• Main Authors: Wang, Yuwen, Han, Mingxiu, Peng, Yudan, Zhao, Ruoqi, Fan, Dongqiong, Meng, Xia, Xu, Hong, Niu, Haijun, Cheng, Jian, Liu, Tao
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.06.2024
• Subjects: Cognitive workload; Deep learning; EEG; Simulated flights; Deep learning; Simulated flights; Cognitive workload; EEG
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2024.106046

•EEG emerges as a promising tool for cognitive workload monitoring.•CNNs are effective in extracting local EEG representations but tend to have limitations in capturing global EEG features.•LGNet aims to simultaneously extract local and global EEG representations.•LGNet employs the SCCE loss function to enhance learning performance. Cognitive workload assessment is crucial for ensuring pilots' safety during flights. Electroencephalography (EEG) is a promising tool for monitoring cognitive workload. Convolutional neural networks (CNNs) are effective in automatically extracting local EEG representations. However, CNNs have limitations in global representations, because a global representation in CNNs normally requires CNNs to be deep enough to have a global receptive field, which normally results in overfitting. To address this issue, we propose a local and global network (LGNet) for assessing two levels of cognitive workload based on EEG during simulated flight. We fuse convolutional and Transformer layers to extract local and global representations from the EEG signals. To enhance the learning performance, we propose a novel SCCE loss, which combines the supervised contrastive loss with the traditional cross-entropy loss. We collect 32-channel EEG data from 10 subjects who perform low and high cognitive workload flight tasks with passive auditory stimuli in a flight simulator on 3 separate days, with each participant performing the task for approximately 345 min. The results show that LGNet with the SCCE loss achieves a 4-fold average classification accuracy of 91.19 % based on cross-clip data partitioning and an average classification accuracy of 83.26 % based on cross-session data partitioning when no target session data is added to the training data. These results significantly outperform classifiers based on handcrafted features and state-of-the-art deep learning methods.