Incremental value of resting electrocardiography and echocardiography for coronary artery disease: a retrospective cross-dataset evaluation

Background: Resting electrocardiography (ECG) and transthoracic echocardiography are widely available in patients evaluated for coronary artery disease (CAD), but their incremental value beyond clinical predictors and their transportability across data sources remain uncertain. Methods: We performed a retrospective modeling study using a curated stenosis-defined angiographic cohort (Z-Alizadeh Sani; n = 303) and a real-world US cohort assembled from MIMIC-IV (n = 8,147). In MIMIC-IV, CAD was operationalized from diagnosis and revascularization records around coronary angiography, and candidate clinical, ECG, and echocardiographic features were extracted within a primary ±30-day peri-angiographic window. Clinical, single-modality, and fusion models were compared using five-fold out-of-fold predictions. We evaluated discrimination, calibration, reclassification, decision-curve utility, subgroup performance, temporal-window sensitivity, and cross-dataset transportability. Results: In the curated cohort, adding ECG and echocardiography to the clinical pretest model increased the AUC from 0.894 to 0.926 (increase, 0.031; P = 0.021) and outperformed the CAD-Consortium basic score (AUC, 0.848; difference, 0.078; P = 0.00015). A broad laboratory panel did not improve discrimination (P = 0.330). In MIMIC-IV, the fusion model achieved an AUC of 0.833 (95% CI, 0.817-0.848), compared with 0.812 (95% CI, 0.795-0.828) for the clinical model (increase, 0.021; P = 5.30 x 10^-7), with a calibration slope of 0.991. The fusion advantage persisted after excluding troponin and across ±7- to ±90-day windows. Cross-dataset transfer was asymmetric: the curated-cohort model decreased from an internal AUC of 0.845 to 0.682 in MIMIC-IV, whereas the MIMIC-IV model decreased from 0.806 to 0.789 in the curated cohort. Conclusions: Resting ECG and echocardiography provided modest but reproducible incremental information beyond clinical predictors in two angiography-related public datasets. However, the MIMIC-IV endpoint was an EHR-derived phenotype and the primary feature window was peri-angiographic rather than strictly pretest. These findings support further pre-angiography validation in lower-prevalence populations before clinical use.