Session S54.3

Automatic Distinguishing between Ischemic and Heart-Rate Related Transient ST-Segment Episodes in Ambulatory ECG Records

J Faganeli*, F Jager

University of Ljubljana
Ljubljana, Slovenia

In ambulatory ECG records ischemia is manifested by transient ST-segment episodes which may or may not be accompanied by an increase in heart rate. There are also transient, heart-rate related, non-ischemic ST-segment episodes present, which are mostly caused by an increase in heart rate. These non-ischemic episodes complicate automatic detection of true ischemic episodes. Our goal was to develop an automatic classification method that distinguishes between this two types of episodes. According to the gold standard established by human expert annotators of the Long-Term ST Database (LTST DB), both types of transient episodes in the records were annotated with regard to changes in heart rate and ST-segment morphology, and accompanied clinical reports. Ischemic ST-segment morphology change class included: horizontal flattening, down sloping, scooping, or elevation; while heart-rate related ST-segment morphology change class included: J-point depression with positive slope, moving of T-wave into ST-segment, T-wave peaking, or parallel shift of ST-segment. In order to automatically classify the episodes we chose the following features: heart rate values, time domain features of ST-segment (deviation, slope and root mean square value) and Legendre orthonormal polynomial coefficients of ST segment. We chose Legendre basis functions because they best fit typical shapes of the ST-segment morphology thus allowing direct insight into the ST-segment morphology changes through the feature space. The first three Legendre basis functions (constant, linear, and square function) are very similar to typical ST-segment morphology changes (level, slope, scooping) during transient episodes. All the features were calculated on 20-second intervals at the beginning and at the extrema of each episode. We then derived, for each feature and for each episode, the difference in feature value at the extrema relative to the beginning of the episode. We evaluated the separability of both types of episodes, given feature, using analysis of variance (ANOVA). We tested the classification method using all records (86) of the LTST DB on all transient ischemic and heart-rate related episodes (1130 ischemic and 234 heart-rate related), as annotated in each single ECG lead, according to the protocol B of the LTST DB. The obtained sensitivity in classifying ischemic versus heart-rate related episodes was 77.8%, while specificity was 73.9%.

(Abstract Control Number: 16)