An innovative intelligent system based on automatic diagnostic feature extraction for diagnosing heart diseases

• Published in: Knowledge-based systems Vol. 75; pp. 224 - 238
• Main Author: Sun, Shuping
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.02.2015
• Subjects: Automation; Boundary curve; Classification; Defects; Diagnosis; Diagnostic systems; Extraction; Heart diseases; Heart sound; Intervals; PCA; SES1 and SES2; Sound; STMHT; STMHT; Heart sound; SES1 and SES2; Boundary curve; PCA
• ISSN: 0950-7051; 1872-7409
• DOI: 10.1016/j.knosys.2014.12.001

•An innovative intelligent system is proposed for diagnosing heart diseases.•A novel procedure is proposed for automatically extracting S1 and S2.•A novel frequency feature matrix (FFM) is defined and automatically extracted.•PCA-based diagnostic features y1 and y2 are generated from the FFM.•SVM-based generated classifiers are used to diagnose heart diseases. An innovative intelligent diagnostic system is proposed in this study, which is primarily reflected in first heart sound (S1) and second heart sound (S2) automatic extraction, frequency feature matrix (FFM) automatic extraction, diagnostic feature y1 and y2 generation based on principal components analysis (PCA) and diagnostic method definition based on the classification boundary curves. Four stages corresponding to the diagnostic system implementation are summarized as follows. Stage 1 describes an envelope ET extraction from heart sound signals. In stage 2, heart sound segmentation points and peaks are first automatically located based on a novel method STMHT, and then S1 and S2 are automatically extracted according to the relationship between the systolic time interval and the diastolic time interval. In stage 3, in the frequency domain, a novel method is first proposed to generate the secondary envelopes SES1 and SES2 for S1 and S2, respectively, and then an STMHT-based FFM is automatically extracted from SES1 and SES2. Finally, the PCA-based diagnostic features y1 and y2 are generated from the FFM. In stage 4, support vector machine (SVM)-based classification curves for the dataset consisting of y1 and y2 are first generated, and then, based on the classification curves, the scatter diagram diagnostic result (SDDR) and numerical diagnostic result (NDR) are defined for diagnosis of heart diseases. The proposed intelligent diagnosis system is validated by sounds from online heart sound databases and by sounds from clinical heart diseases. As a result, the classification accuracies (CA) achieved are 91.7%, 98.8%, 98.4%, 99.8%, 98.7%, 97.8%, 98.1% and 96.5% for the detection of atrial fibrillation (AF), aortic regurgitation(AR), mitral regurgitation (MR), normal sound (NM), pulmonary stenosis (PS), small ventricular septal defect (SVSD), medium ventricular septal defect (MVSD) and large ventricular septal defect (LVSD), respectively.