Clinical evaluation of algorithms for ST measurement during exercise test

Computer processing of the exercise electrocardiogram (ECG) has many advantages, but the reliability of the analysis algorithms is not easily evaluable. No standard annotated database, nor recommended practice for testing and reporting performance results is available: thus, performance evaluation of such devices can be accomplished only by using a set of unannotated recordings, obtained in clinical practice. We evaluated the accuracy of an original microcomputer-based exercise test analyzer comparing the ST computer output with the measurements obtained by two experienced cardiologists.

Six hundred ECG strips were randomly selected from the exercise test recordings of 60 patients. The ST shift (at J + 80 ms) was blindly assessed by two observers (with the aid of a calibrated lens) and compared with computer measurements. Correlation coefficients, linear regression equations, percent of discrepant measurements, and 95% confidence limits of the mean error were calculated for all leads, peripheral leads, precordial leads, and "stress-test" leads (II, III, aVF, V4, V5, V6).

The computer did not analyze five samples on a total of 600 (0.83%) ECG strips because of excessive noise or signal loss, while 51 (8.5%) were considered unreadable by both observers and 67 (11.2%) were rejected by at least one observer. Correlation between the measurements taken by computer and observer(s) measurements was statistically significant (p < 0.001 for all lead groups), no systematic measurement bias was found, and the mean difference was lower than human eye resolution.

Our algorithms provide results as good as those provided by trained cardiologists in measuring ST changes occurring during exercise test. However, this study did not evaluate whether computer improvement of the signal-to-noise ratio would allow accurate measurements even on cardiologists' uninterpretable ECG. This potential advantage of computer-assisted analysis could be assessed only by using a dedicated exercise test database, in which different patterns of noise are superimposed on noise-free recordings previously annotated for ST level.