COMPUTERS AND BIOMEDICAL RESEARCH 20, 410-427 (1987) ECG Waveform Analysis by Significant Point Extraction I. Data Reduction Ho Soo LEE,* QUIN-LAN CHENG,~ AND NITISH V. THAKOR? * IBM Thomas .I. Watson Research Center. Yorktown Hei,yhts. Rie~, 1’orX 10549; and f Department of Biomedical Engineering. Johns Hopkins University. Baltimore, Maryland 21205 Received June 13. 1986 We present a new technique for automatic data reduction and pattern recognition of time- domain signals such as electrocardiogram (ECG) waveforms. Data reduction is important because only a few significant features of each heart beat are of interest in pattern analysis. while the patient data collection system acquires an enormous number of data samples. We present a significant point extraction algorithm, based on the analysis of curvature, that identifies data samples that represent clinically significant information in the ECG wave- form. Data reduction rates of up to I : 10 are possible without significantly distorting the appearance of the waveform. This method is unique in that common procedures help in both data reduction as well as pattern recognition. Part II of this work deals specifically with pattern analysis of normal and abnormal heart beats. c~ 1987 Academic POW. IX 1. INTRODUCTION Analysis of electrocardiogram (ECG) signals by automatic patient monitor- ing instruments is of great medical interest. Automated instruments now con- tinuously monitor patients in intensive care units and generate alarms if abnor- mal ECG patterns are observed (I). In the course of patient monitoring more than 100,000 ECG beats sampled at a rate of about 250 samples/set must be analyzed every day. In fact, only a few significant features are usually of interest: such as, beat-to-beat interval, height, duration, and polarity of each beat (2). Data reduction algorithms are needed to retain only selected samples that adequately describe the ECG pattern and thus reduce the need for memory and processing. Data reduction is usually followed by automated pattern analy- sis to identify abnormal beats and thus related heart disease. Several data reduction and feature extraction algorithms have been reported in the literature. Rosenfeld and Johnson (3) and Rosenfeld and Weszka (4) analyzed chain codes of patterns. Curvatures, based on calculations of angles and points of inflection, were determined. Davis (5) extended this to hierarchi- cal definitions of graphs. Freeman and Davis (6), and Sankar and Sharma (7) similarly identified corners or dominant points from the analysis of closed curves. Analysis of continuous time-domain waveforms, rather than images 410 OOIO-4809/87 $3.00 Copyright 0 1987 by Academic Press. Inc. All rights of reproduction in any form reserved.