A novel bidirectional LSTM network based on scale factor for atrial fibrillation signals classification. (July 2022)

• Record Type: Journal Article
• Authors: Feng, Kunye; Fan, Zile
• Is Part Of: Biomedical signal processing and control. Volume 76(2022)
• Publication Date: 2022-07
• Subjects: Atrial fibrillation -- Long short term memory -- Scale factor; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103663
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: We come up with IB-LSTM network by introducing a scale factor into existing B-LSTM for AF detection. Taking physiological and pathological characteristics into account, our IB-LSTM has certain model interpretability. Our work offers the first empirical exploration of redesign architecture of B-LSTM. Abstract: Atrial fibrillation (AF) is a common arrhythmia worldwide. The visual examination for electrocardiogram is the major diagnosis method, which is generally burdensome and inefficient. In this work, an improved bidirectional long short term memory (IB-LSTM) frame is specially designed for intelligence-based AF signals classification. Based on the existing B-LSTM architecture, a scale factor is installed into IB-LSTM network that enables the model system to effectively reallocate information representation and therefore learn better representation. After several pre-processing steps such as denoising, these results demonstrate consistent performance improvements with the accuracy of 98.2% and 97.5% against multiple existing approaches and lower computation costs than B-LSTM with the two public MIT-BIH AF and arrhythmia databases. The proposed IB-LSTM network is capable of trading off between model accuracy and computing resources. In particular, this research provides the first empirical exploration of redesign architecture of B-LSTM to alleviate high computation cost and information redundancy, which shows great prospects as efficient auxiliary tools for medical diagnosis.