Response to letter by Drs Dilaveris and Stefanadis: Electrocardiographic predictors of atrial fibrillation: Methodological considerations

We would like to thank Drs Dilaveris and Stefanidis for their interest in and insightful comments regarding our recent study on the electrocardiographic (ECG) predictors of atrial fibrillation (AF).1 In their letter, they noted concerns regarding the automation of the ECG measurements without manual verification. We agree with the points raised and understand the limitations of computerized measurements. Although some automated measurements such as the QT interval have been shown to be reproducible,2 others such as degree of ST-segment deviation have been shown to be less accurate.3

Although computerized measurements may not be as accurate as manual measurements, they did make feasible a study of this size, which included >42,000 ECGs. Computerized errors will lead to misclassification; however, because the measurements were not influenced by future development of AF, the misclassification was nondifferential. Nondifferential misclassification bias results in dilution of effect size, meaning that our reported hazard ratios were underestimates of the true association. The advantage however is that, with a large sample size, we have much greater power to identify statistically significant associations and to measure the independent nature of these effects. At the level of the individual, however, validation of the computerized measurements such as P-wave dispersion and index will have to be made before automated measurements are used clinically.

Drs Dilaveris and Stefanidis also comment on the inherent limitations of the use of a single 12-lead ECG recording that include presence of artifact, lack of distinct P-wave onset and offset, and inability to capture physiologic variation over time. Manual onscreen verification of measurements, as suggested,4 would indeed help improve accuracy. Technological advances, such as signal averaging as mentioned, as well as improved sensitivity and signal processing will also help overcome some of these limitations. Clinically, however, the existing 12-lead ECG system is the most commonly used test to evaluate cardiac patients; and tools for improving the accuracy of P-wave measurements are not yet widely available. As interest for the prediction of AF grows, we hope that these tools become more commonplace.