Sources of Variability In Qt Calculations

Background: In the 2006 Computers and Cardiology Challenge, the PTB diagnostic ECG database was used for QT interval measurement. There were 549 records each 6 to 7 minutes long rich in well annotated abnormal records but also containing 80 normal control records. Division 1 of the QT Interval Measurement Challenge had fifteen participants who “made their measurements manually or using semi-automated methods that allowed for manual correction or adjustment”. These results were combined to create a gold standard for scoring the other divisions of the QT Challenge. Included in the developed gold standard was the mean absolute difference between the individual measurements and the reference medians, in milliseconds. This information can be used to infer the difficulty that the reviewers encountered in measuring lead II of the 15 lead data (12 standard leads plus Frank leads). In turn, when combined with feature extraction from the ECGs in the database, can lead to insights on the situations where measurement is most difficult. Methods: Philips DXL algorithm was used to create multiple measurement table results to characterize each record. Each record was analyzed in up to twelve 10 s segments and the results combined into average values for each record. Spearman rank-order correlation values (rho) were calculated between the MAD variability values for each record and standard ECG measurements like T-wave axis and T-wave amplitude in lead II. In some cases the same analysis was performed on the absolute value of the measurement. Results: Numerous relatively large correlations were seen when the reviewers differed. Over the full database high variability correlated well with small T amplitude in lead II, negative T amplitude in lead II, low absolute T amplitude over all leads, low T area in lead II and biphasic T-waves. Variability also correlated with T-axis in both horizontal and frontal planes. Negligible correlations were seen with heart rate, QRS duration and QT value. The mean reviewer variability over all records was 18 ms but only 12 ms on normal controls. On the normal controls the only features with small correlation are low T area in lead II and biphasic T-waves. Conclusions: Abnormal ECGs are significantly more difficult to read than normal controls. Certain T-wave features, when present, tend to increase the variability between even expert readers. Lead II features showed more correlation with high variability than any other lead. Developers should