CosELU: An improved activation deep neural network for EEG signal recognition in stroke patients

Shuaishi Liu; Wenkai Ji; Qiye Zhou; Xiaoqin Duan; Jiang Yi; Changxian Xu; Zaixiang Pang; Zhongbo Sun

doi:10.1016/j.neunet.2025.107800

CosELU: An improved activation deep neural network for EEG signal recognition in stroke patients

Neural Netw. 2025 Jul 5:191:107800. doi: 10.1016/j.neunet.2025.107800. Online ahead of print.

Authors

Shuaishi Liu¹, Wenkai Ji², Qiye Zhou³, Xiaoqin Duan⁴, Jiang Yi⁵, Changxian Xu⁶, Zaixiang Pang⁷, Zhongbo Sun⁸

Affiliations

¹ College of Electrical and Electronic Engineering, Changchun University of Technology, Changchun, 130012, China. Electronic address: liushuaishi@mail.ccut.edu.cn.
² College of Electrical and Electronic Engineering, Changchun University of Technology, Changchun, 130012, China.
³ Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, 999077, Hong Kong. Electronic address: 1155191415@link.cuhk.edu.hk.
⁴ The 2nd Hospital of JiLin University, Changchun, 130041, China. Electronic address: duanxq@jlu.edu.cn.
⁵ The 2nd Hospital of JiLin University, Changchun, 130041, China. Electronic address: qiyi328@jlu.edu.cn.
⁶ College of Mechanical and Electrical Engineering, Changchun University of Technology, Changchun, 130012, China. Electronic address: 2202004073@stu.ccut.edu.cn.
⁷ College of Mechanical and Electrical Engineering, Changchun University of Technology, Changchun, 130012, China. Electronic address: pangzaixiang@ccut.edu.cn.
⁸ College of Electrical and Electronic Engineering, Changchun University of Technology, Changchun, 130012, China. Electronic address: zbsun@ccut.edu.cn.

PMID: 40633296
DOI: 10.1016/j.neunet.2025.107800

Abstract

Motor imagery electroencephalograme (MI-EEG) signal recognition algorithms are essential for rehabilitation training of clinical stroke patients. However, EEG signals of stroke patients using traditional activation functions are prone to exploding gradient and vanishing gradient problems in deep learning deep neural network training, leading to a decrease in the recognition rate of EEG signals. In this paper, an adaptive activation function is developed and investigated to address this problem. The method monitors the gradient of the activation function by adjustable parameter a and adaptive parameter λ_i⁺, which alleviates the exploding gradient and vanishing gradient problems. In addition, a suitable attention mechanism convolutional neural network (AMCNN) is designed for the classification of EEG signals, and CosELU is utilized to the AMCNN network for training of EEG signals, and finally the effectiveness of the CosELU activation function is verified on a stroke patient dataset. The source code is publicly available at https://github.com/ccut3329lab/jwk_code.git.

Keywords: AMCNN; CosELU; Exploding gradient problem; MI-EEG; Vanishing gradient problem.