QRS morphology representation and noise estimation using the Karhunen-Loeve transform

The authors describe robust methods for deriving Karhunen-Loeve (KL) basis functions, which can be used to represent the QRS complex. Using a five-term KL expansion of a 200-ms interval, which includes the QRS complex and part of the ST segment, one can represent morphology on two simultaneous ECG (electrocardiographic) leads with sufficient fidelity for beat classification. The residual error of the representation is an ideal estimate of the instantaneous noise content of the signal and permits identification of events for which the morphologic information is unreliable. The authors have compared the performance of the current KL-based arrhythmia analysis program with its predecessor (which uses a set of time-domain features for morphology representation but is otherwise identical to the newer program). In evaluations using the MIT-BIH and AHA (American Heart Association) databases, and a newly developed database containing approximately 2.5 million annotated beats (including over 80000 premature ventricular contractions) from 27 long-term ECG recordings, it was found that beat classification errors using the KL transform were as little as one-fourth of those for the older program.<>