An attention-based CNN-BiLSTM hybrid neural network enhanced with features of discrete wavelet transformation for fetal acidosis classification. (30th December 2021)

• Record Type: Journal Article
• Authors: Liu, Mujun; Lu, Yaosheng; Long, Shun; Bai, Jieyun; Lian, Wanmin
• Is Part Of: Expert systems with applications. Volume 186(2021)
• Publication Date: 2021-12-30
• Subjects: Cardiotocography monitoring -- Fetal acidosis -- Deep learning -- Discrete wavelet transformation features -- Sigmoid classification; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2021.115714
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: CNN-BiLSTM can integrate spatial features and temporal correlation of FHR signals. Attention is focused on more relevant features for fetal acidosis classification. DWT can reduce the overfitting and improve the accuracy of model classification. Abstract: Cardiotocography (CTG) is widely used in fetal monitoring, especially in the diagnosis of fetal acidosis. However, the manual interpretation of CTG analysis may easily lead to a low diagnostic rate, usually caused by various subjective factors. In order to reduce misdiagnosis, we propose an attention-based CNN-BiLSTM hybrid neural network enhanced with features of discrete wavelet transformation (DWT) for fetal acidosis classification. A joint model of convolutional neural network (CNN) and bi-directional long short-term memory (BiLSTM) is established to capture the complex nonlinear spatial and temporal relations of fetal heart rate (FHR) signals. The attention mechanism is then adopted to focus on important input features. And DWT is used to obtain FHR signals transformation coefficient features in order to reduce overfitting. Two features are fused together to classify fetal acidosis. This study uses signals from the public databases of the CTU-UHB for evaluation. A ten different verifications yields average sensitivity (SE), specificity (SP), and quality index (QI) of 75.23%, 70.82% and 72.29%, respectively. Our approach achieves better experimental results than previous works. Moreover, Our hybrid model is an end to end one, with a much simpler DWT feature extraction. With the advent of the big data era, our hybrid model will have great advantages.