Cross-subject Drowsiness Recognition Model Based on the Multi-scale Parallel Convolution of EEG Channels

Cross-subject electroencephalogram (EEG) drowsiness recognition is currently one of the most efficient methods. However, the traditional cross-subject approaches overlook the correlation between channel sub-features and effectively utilize the potential connection between frequency features and drow...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:2023 IEEE Intl Conf on Parallel & Distributed Processing with Applications, Big Data & Cloud Computing, Sustainable Computing & Communications, Social Computing & Networking (ISPA/BDCloud/SocialCom/SustainCom) s. 709 - 714
Hlavní autoři: Li, Songquan, Zhu, Rongbo, Yuan, Jun, Wan, Shaohua
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 21.12.2023
Témata:
Brain modeling
Convolution
Correlation
Cross-subject
Deep learning
Drowsiness recognition
Electric potential
Electroencephalography
Electroencephalography (EEG)
Feature extraction
Grouped features
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Cross-subject electroencephalogram (EEG) drowsiness recognition is currently one of the most efficient methods. However, the traditional cross-subject approaches overlook the correlation between channel sub-features and effectively utilize the potential connection between frequency features and drowsiness, making it difficult to extract shared features from differentiated subjects. This paper proposes a novel deep learning model called CPCNN to address the challenging of cross-subject drowsiness recognition without calibration and improve accuracy. CPCNN obtains delta frequency band features related to drowsiness from sub-features in the channel dimension, effectively combining the advantages of delta frequency bands in EEG with the richness of channel features. To capture the common features of drowsiness, a channel slicing module is proposed, which divides the spatial and channel features extracted by the deep separable convolution module into multiple sub-features along the channel dimension. In addition, to further improve cross-subject recognition performance, a multi-scale parallel group convolution is designed to capture four key frequency sub-feature groups within the delta frequency band. The experimental results show that CPCNN boosts the recognition performance, significantly outperforms all strong baselines. On balanced and unbalanced public datasets, the average accuracy of CPCNN is 81.49% and 78.26%, respectively, surpassing the excellent baseline method by 3.14% and 0.56%.
DOI:10.1109/ISPA-BDCloud-SocialCom-SustainCom59178.2023.00125