Table shows baseline characteristics of the Disease and Healthy groups and of the two Disease subgroups in the training set. Besides being free of hypertension and diabetes and having lower body mass index and medication use, the Healthy group training set also was younger than the Disease group training set. Because of the age disparity, two sets of primary A-ECG scores were constructed and validated in the training set for later evaluation in the test set: one wherein all healthy subjects were included and one wherein only those healthy subjects >40 years of age (mean 51 ± 8 years, N = 133, 63% men) were included. For the additional 315 individuals who comprised the test set, the distributions of hypertension, diabetes, body mass index, LVEF, and medication use were similar to those shown in Table . The mean ages in the test set were 59 ± 12, 59 ± 13, 56 ± 12 and 49 ± 11 years for the Disease group, the with-LVSD and without-LVSD Disease subgroups and the Healthy group respectively, men comprising 65%, 74%, 60% and 57% of those groups, respectively.
Demographic Characteristics of the Training Set Disease Group, Disease Subgroups and Healthy Group
Figures and show how the performance of an A-ECG score in the training set depended on the number of ECG parameters the score incorporated. For primary ("Healthy versus Disease") A-ECG scores (Figure , N = 708), only negligible further gains in cross-validated accuracy occurred with scores containing more than ~9 parameters. For secondary ("Disease with versus without LVSD") A-ECG scores (Figure , N = 290), this same cutoff occurred at only ~5 parameters. The first and second parameters incorporated into primary A-ECG scores by the automatic selection procedures were the QTVI in lead II and the spatial mean QRS-T angle, respectively (Figure ). For secondary (LVSD) A-ECG scores, the first and second parameters incorporated by the same procedures were the Z integral and the spatial mean QRS-T angle, respectively (Figure ).
Effect of number of parameters in a primary ("Healthy versus Disease") Advanced ECG (A-ECG) score on the score's jackknifed accuracy in the training set (N = 708).
Effect of the number of parameters in a secondary ("Disease with versus without left ventricular systolic dysfunction", LVSD) Advanced ECG (A-ECG) score on the score's jackknifed accuracy in the training set (N = 290).
Table shows the performances in the training set of the pooled, strictly conventional ECG criteria, along with those of the most relevant single parameters and A-ECG scores. The candidate conventional ECG criteria outlined in the Methods section were retrospectively optimized when their Sokolow-Lyon subcriteria were dropped and replaced instead by subcriteria for left atrial abnormality (P-wave duration >120 ms or terminal negative component of a biphasic P-wave in lead V1 >4 ms*mV in area). Thus, only the resulting optimized set of conventional ECG criteria was carried forward for later use with the test set. Not unexpectedly, the retrospectively optimized A-ECG scores outperformed the retrospectively optimized pooled conventional ECG criteria in the training set. Of note, the optimal primary A-ECG scores made use of the entire ~5-min (so called "full-disclosure") 12-lead recording because they incorporated results from QTVI (Figure ). Inasmuch as most 12-lead ECG machines do not yet have full-disclosure capabilities, Table also shows the diagnostic performance in the training set of an optimized primary A-ECG score that was only allowed to incorporate results from parameters likely yielding reliable and reproducible results within strictly "snapshot" (10-sec) ECG recordings.
Accuracies and Predictive Values of Pooled Conventional versus A-ECG Criteria in the Training Set
Table shows the performances in the test set
of the optimized pooled conventional ECG criteria and of the most relevant single parameters and A-ECG scores generated in the training set. Although as expected most A-ECG scores tended to have slightly diminished performance in the test set compared to the training set (compare Table to Table ), several primary A-ECG scores generated from the training set still had accuracies of 90% or greater in the test set. For example, compared to the optimized pooled criteria from the strictly conventional ECG, the best 7-parameter primary full-disclosure A-ECG score generated in the training set increased the sensitivity of resting ECG for identifying Disease in the test set from 78% (72-84%) to 92% (88-96%) (P < 0.0001) while also increasing specificity from 85% (77-91%) to 94% (88-98%) (P < 0.05). Another 7-parameter A-ECG score that only incorporated parameters likely yielding reliable and reproducible results within "snapshot" ECG recordings was only slightly less accurate. In diseased patients, another 5-parameter secondary
A-ECG score generated in the training set also increased the PPV of ECG for additionally predicting LVSD in the test set from 53% (41-65%) to 92% (78-98%) (P < 0.0001) without significantly compromising NPV. This secondary A-ECG score had corresponding positive and negative likelihood ratios for LVSD in the test set of 12.16 and 0.18, respectively, versus 1.23 and 0.21 for the optimized pooled conventional ECG criteria. The exact components and coefficients of those training set-generated primary and secondary A-ECG scores that performed best in the test set are shown in Additional file 2
(Supplemental Table 2).
Accuracies and Predictive Values of Pooled Conventional versus A-ECG Criteria in the Test Set