Quantifying and disentangling sources of QRS duration and QTc interval divergence between Korean (SNUH) and US (EchoNext) wide QRS cohorts

BACKGROUND AND OBJECTIVES AI-based ECG interpretation risks miscalibration when models trained on one national cohort are applied to another. This study quantified the QRS duration (QRSd) and corrected QT interval (QTc) gap between Korean and US cohorts, controlling for demographic confounders, and tested whether this gap reflects ethnicity specifically or broader country/site-level factors. METHODS This retrospective study used Korean SNUH (n = 45,249) and US EchoNext (n = 12,993; Wide QRS subset, QRSd ≥120 ms) cohorts, matched for clinical comparability via propensity score sensitivity analysis. Mean differences were assessed by Welch's t-test; the independent ethnicity effect was quantified via ANCOVA (QRSd or QTc ∼ Ethnicity + Age + Sex). A race/ethnicity subgroup analysis within EchoNext (White, Black, Hispanic, Asian) tested whether the gap reflected racial composition. RESULTS Both parameters differed significantly (p < 0.0001). US QRSd (140.59 ms) exceeded Korean QRSd (112.66 ms; difference -27.93 ms). Ethnicity/cohort membership remained strongly associated with QRSd and QTc after adjustment for age and sex: β = -27.52 ms for QRSd, β = -26.26 ms for QTc (both p < 0.0001). The gap remained similar in magnitude (-26 to -28 ms) across all EchoNext race/ethnicity subgroups. CONCLUSIONS A clinically meaningful 25-27 ms QRSd/QTc gap persists between Korean and US cohorts. Its stability across EchoNext's own race/ethnicity subgroups indicates country/site-level factors, not ethnicity alone, likely contribute. This gap nonetheless represents a real, reproducible technical risk of misclassification when ECG AI models trained on one national cohort are applied to another, supporting cohort-diverse validation regardless of causal mechanism.