Cumulative electrocardiogram abnormalities predict left ventricular ejection fraction reduction: EchoNext database analysis.

BACKGROUND

Individual electrocardiographic abnormalities associate with reduced left ventricular ejection fraction (LVEF), but whether multiple continuous ECG parameters provide synergistic predictive value compared to single parameters or binary scoring remains incompletely characterized.

METHODS

QRS duration, PR interval, and heart rate-corrected QT interval (QTc) from ECGs were analyzed in 89,630 paired ECG-echocardiogram studies from the EchoNext database. After systematic evaluation of missing data patterns revealing significant selection bias, the cohort was divided into training (70%, n = 62,741) and validation (30%, n = 26,889) sets using stratified sampling. Seven logistic regression models were developed, all adjusted for age and sex. Models were evaluated using bootstrap confidence intervals with 1000 iterations and 5-fold stratified cross-validation.

RESULTS

The full continuous model achieved fair discrimination (AUC 0.710, 95% CI 0.699-0.714) compared to QTc alone (AUC 0.695, ΔAUC = 0.016, p < 0.001). Among individual parameters, QTc demonstrated strongest association (AUC 0.695), followed by QRS duration (AUC 0.673) and PR interval (AUC 0.603). LVEF ≤45% prevalence increased from 15.9% (0 abnormalities) to 30.6% (1 abnormality), 49.5% (2 abnormalities), and 58.3% (3 abnormalities) (Cochran-Armitage Z = 76.0, p < 0.001). At the optimal threshold, the model achieved sensitivity 59.9%, specificity 72.6%, positive predictive value 37.9%, and negative predictive value 86.7%. All models showed excellent calibration slopes (range 0.963-1.043). Cross-validation confirmed stability (CV AUC 0.706 ± 0.006).

CONCLUSIONS

Combined continuous ECG parameters provide predictive value for reduced LVEF compared to individual parameters or binary classifications. However, selection bias from non-random missing data (42.0% vs 21.8% outcome prevalence in incomplete vs complete cases, p < 0.001) suggests findings likely underestimate true associations.