CraftNet: A deep learning ensemble to diagnose cardiovascular diseases. (September 2020)

• Record Type: Journal Article
• Authors: Li, Yong; He, Zihang; Wang, Heng; Li, Bohan; Li, Fengnan; Gao, Ying; Ye, Xiang
• Is Part Of: Biomedical signal processing and control. Volume 62(2020)
• Publication Date: 2020-09
• Subjects: Cardiovascular diseases -- Deep learning -- Handcraft features; 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.2020.102091
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Combining handcraft features with neural networks to alleviate the data imbalance in the heart disease diagnosis. Assembling multiple neural networks in a model. Devising a novel network structure and a loss function to further empower our model. Outperforming the state-of-the-art four-category classification methods on the public MITBIH dataset. Abstract: The early diagnose of cardiovascular diseases (CVDs) is important and has attracted a lot of research attention. It can rescue more than 17 million people a year or alleviate their symptoms. Research interest has been devoted to the handcraft features and deep features for diagnosing CVDs from electrocardiograph (ECG). However, existing classifiers on handcraft features lacked the robust classification ability, while the deep neural networks are strongly affected by data imbalance. This paper proposed designing a simple architecture of deep neural network, CraftNet, for accurately recognizing the handcraft features. It assembled multiple child classifiers according to decision directed acyclic graph. The classifiers have a tailored structure for classifying the handcraft features, with a mixed loss function, named P-S loss, to optimize it. CraftNet has the advantages of both handcraft features and deep learning methods, i.e., it has a stronger classification ability and is less affected by data imbalance. The proposed CraftNet was tested on the public MIT-BIH dataset. Experimental results showed that it achieved the sensitivity 88.16%, 85.37%, 94.53%, and 88.92% for four categories, and increased the average sensitive accuracy from 86.82% to 89.25%, verifying the robust recognition ability of CraftNet.