Acute coronary occlusion in chest pain patients with broad QRS complex – artificial intelligence-enhanced ECG interpretation compared to conventional electrocardiographic criteria

Overview

Recognising acute coronary occlusion (ACO) on the ECG is especially hard when the QRS complex is broad: criteria for left bundle branch block (LBBB) and ventricular pacing are complex, and none exist for right bundle branch block (RBBB) or non-specific intraventricular conduction delay (NIVCD). This independent single-centre study from the University of Cologne tested an AI-enhanced ECG model in 439 consecutive chest pain patients with QRS ≥120 ms, of whom 14 (3.2%) had ACO. The AI model detected 8 of 14 occlusions with 98.1% specificity and an AUC of 0.92, and kept 98.4% specificity in RBBB, where no ECG reference standard exists. In LBBB it performed similarly to the Smith-modified Sgarbossa and Barcelona criteria, and in ventricular pacing it classified all patients correctly (PPV 100% vs 9.1–20.0% for conventional criteria), although the number of occlusions was small.

Published in: ESC Congress 2026 (ePoster)
Published on: 30 August 2026

Background

Identifying acute coronary occlusion (ACO) in chest pain patients with wide QRS complexes remains challenging. Electrocardiographic criteria exist for left bundle branch block (LBBB) and ventricular pacing (VP), but their complexity limits routine use, and no standardised ECG-based framework is available for right bundle branch block (RBBB) or non-specific intraventricular conduction delay (NIVCD). Artificial intelligence (AI)-enhanced ECG analysis may add value, but its performance in these subgroups is unclear. This study evaluated the diagnostic performance of an AI-enhanced ECG model for detecting ACO in chest pain patients with bundle branch block or NIVCD in a metropolitan chest pain unit, and compared it with established ECG criteria in LBBB and VP.

Methods

The AI model, a deep convolutional neural network, was previously trained and externally validated in international cohorts. All consecutive chest pain patients presenting to the chest pain unit of the University of Cologne, Germany, between November 2018 and June 2021 with a QRS duration ≥120 ms were included. ECGs were independently classified by two investigators, with adjudication by a third if needed, and clinical management followed European Society of Cardiology guidelines. The endpoint was ACO, adjudicated by an independent, blinded physician using coronary angiography, biomarkers and echocardiography. Diagnostic performance was assessed by receiver operating characteristic (ROC) analysis. In LBBB and VP, the AI model was compared with the Sgarbossa, Smith-modified Sgarbossa and Barcelona criteria.

Results

Of 439 patients (median QRS 144 ms [132–158]), 192 had RBBB (43.7%), 113 LBBB (25.7%), 70 NIVCD (15.9%) and 64 VP (14.6%). ACO prevalence was 3.2% (14 patients: 9 RBBB, 4 LBBB, 1 VP). Overall, the AI model detected 8 of 14 ACO cases (sensitivity 57.1%) with an AUC of 0.92, specificity of 98.1%, accuracy of 96.8%, PPV of 50.0% and NPV of 98.6%, and produced two false positives in the low-risk subgroup (n = 216). In RBBB, the model reached an AUC of 0.93, specificity of 98.4% and sensitivity of 44.4%; in NIVCD, where no ACO occurred, specificity was 98.6%. In VP, the AI model classified all patients correctly (AUC 1.00, PPV 100%), whereas the Sgarbossa, Smith-modified Sgarbossa and Barcelona criteria detected the single ACO case but had PPVs of 20.0%, 14.3% and 9.1%. In LBBB, the AI model (AUC 0.88; sensitivity 75.0%, specificity 96.3%) performed similarly to the Smith-modified Sgarbossa (AUC 0.87; 75.0%, 99.1%) and Barcelona criteria (AUC 0.92; 100%, 84.4%), and better than the original Sgarbossa criteria (AUC 0.62; sensitivity 25.0%).

Conclusion

The AI-enhanced ECG model showed high specificity in patients with RBBB, a subgroup with no established ECG reference standard, which warrants further evaluation in independent cohorts. In LBBB, it achieved diagnostic performance comparable to established but complex ECG criteria, and in ventricular pacing it showed superior classification accuracy. Subject to external validation in cohorts with a higher prevalence of ACO, this approach may offer a rapid and pragmatic diagnostic alternative.