.png){kind=icon}
.png){kind=icon}
Return
Multiple EEG chanel attention based multi-task model for epileptiform activity quantification
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.132342.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Epileptiform activity is critical in characterizing electroencephalogram (EEG) on electrical status epilepticus during sleep (ESES). Spike-wave index (SWI) is a quantitative measure used to characterize the intensity of epileptiform activity, and it is helpful for the clinical diagnosis of ESES encephalopathy. The core of SWI calculation lies in two tasks: sleep stage determination and epileptiform activity segmentation. The current research usually focuses only on single task while the inner-connection between the two tasks is not well exploited. To address this issue, a multi-task learning method with EEG channel attention module for epileptiform activity quantification analysis (MCAQN) is developed in this paper. It consists of a task-sharing network and two task-specific attention networks. The task-sharing network is responsible for learning the global features of EEG. And two task-specific attention networks separately focus on the task-specific information extracted from the global features by EEG channel attention module (ECA), with one dedicated to sleep staging and SWI quantification. To balance the importance of multi-task, a novel loss function is designed that combining semantic segmentation based SWI quantization loss and sleep stage classification loss. This multi-task learning approach allows each module to capture task-specific information while simultaneously leveraging the inherent correlations among related tasks, thereby substantially improving the overall performance. Experiment was conducted on the long-term EEG data from the Children’s Hospital of Zhejiang University School of Medicine (CHZU) dataset. The proposed MCAQN has demonstrated obvious advantages over single-task methods and some state-of-the-art methods (SOTA) methods. Specifically, it can quantify SWI accurately without relying on thresholds and expert experience, and the SWI error ( ) is low to 2.406 %.
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W
View more organizations