Application of multi-feature fusion and random forests to the automated detection of myocardial infarction. (January 2020)

• Record Type: Journal Article
• Authors: Wang, Zhizhong; Qian, Longlong; Han, Chuang; Shi, Li
• Is Part Of: Cognitive systems research. Volume 59(2020)
• Page Start: 15
• Page End: 26
• Publication Date: 2020-01
• Subjects: MI -- ECG -- Principal component analysis -- Statistical feature calculation -- Entropy calculation -- Random forests; Cognition -- Periodicals; Cognitive engineering (System design) -- Periodicals; Artificial intelligence -- Periodicals; 153.05
• DOI: 10.1016/j.cogsys.2019.09.001
• Languages: English
• ISSNs: 1389-0417

Abstract

Highlights: A novel two-layer framework along with two-layer features is proposed. The method performs vital clinical significance based on inter-patient schemes. The results shows high classification performance using random forests. Abstract: Myocardial infarction (MI) was one of the most threatening cardiovascular diseases due to its suddenness and high mortality. Electrocardiography (ECG) reflected the electrophysiological activity of the heart which was widely used for the diagnosis of MI. The aim of the paper was to provide a novel method to detect MI leveraging ECG. Firstly, data enhancement technology was employed to extend the database and prevent overfitting. Then, principal component analysis (PCA) features, statistical features, and entropy features were computed as the representation of first layer features for each lead. Furthermore, the second layer features for each lead were extracted by using random forests (RF), and the feature extraction results were quantified as a classification data set. Finally, in order to evaluate the proposed method, two schemes for the intra-patient and inter-patient were employed. The accuracy, sensitivity, specificity and F 1 values in the intra-patient scheme were 99.71%, 99.7%, 99.73%, and 99.71%, respectively, and 85.82%, 73.91%, 97.73%, and 83.9% in the inter-patient scheme. Meanwhile, compared with different methods including support vector machine (SVM), back propagation neural network (BPNN), and k-nearest neighbor (KNN), RF displayed the best performance.