3.1 Introduction

Chapter 1 presented a description of the ECG in terms of its etiology and clinical features, and Chapter 2 an overview of the possible sources of error introduced in the hardware collection and data archiving stages. With this groundwork in mind, this chapter is intended to introduce the reader to the ECG using a signal processing approach. The ECG typically exhibits both persistent features (such as the average P-QRS-T morphology and the short-term average heart rate, or average RR interval), and nonstationary features (such as the individual RR and QT intervals, and long-term heart rate trends). Since changes in the ECG are quasi-periodic (on a beat-to-beat, daily, and perhaps even monthly basis), the frequency can be quantified in both statistical terms (mean, variance) and via spectral estimation methods. In essence, all these statistics quantify the power or degree to which an oscillation is present in a particular frequency band (or at a particular scale), often expressed as a ratio to power in another band. Even for scale-free approaches (such as wavelets), the process of feature extraction tends to have a bias for a particular scale which is appropriate for the particular data set being analyzed. ECG statistics can be evaluated directly on the ECG signal, or on features extracted from the ECG. The latter category can be broken down into either morphology-based features (such as ST level) or timing-based statistics (such as heart rate variability). Before discussing these derived statistics, an overview of the ECG...