Automatic EEG artifact detection using a local-global feature fusion network in time and time-frequency domains

Video electroencephalographies (VEEGs) are often affected by artifacts, which can diminish clinicians' efficiency in interpreting VEEG data and potentially result in diagnostic errors. Conversely, certain ictal VEEG artifacts caused by automatisms, such as rhythmic chewing and blinking, may offer significant diagnostic clues for temporal lobe epilepsy. To address the challenges mentioned above, this study aims to develop an algorithm capable of automatically identifying and classifying artifact types, assisting clinicians in distinguishing interference signals from seizure signals.

This paper proposes a Local-Global Feature Fusion Network based on Time-Domain and Time-Frequency Domain (LG-TDTFD-Net) for detecting and classifying temporal lobe epilepsy artifacts. The model processes time-domain and time-frequency signals separately, leveraging convolutional neural network (CNN) to extract local features from both domains. Additionally, convolution operations are performed on the features extracted by the Transformer to capture deep local features, resulting in a feature set that simultaneously encompasses global, deep, and shallow local information.

Experimental results demonstrate that the proposed method excels in detecting and classifying temporal lobe epilepsy artifacts, outperforming existing baseline models on clinical and public datasets.

The method effectively exploits the complementarity among the three feature types by utilizing feature concatenation and CNN-based fusion, enhancing feature representation and improving model generalization.

The Chinese Clinical Trial Registry, ChiCTR2300079319. Registered 30 December 2023, https://www.chictr.org.cn/showproj.html?proj=214123 .