Multicentre validation of an AI model for detection of reduced left ventricular ejection fraction
R S Rama, J M Lillo Castellano, R Lopez Garcia, S Manzano Fernandez, J A Gomez Company, L Alvarez Acosta, A de la Linde Valdes, I Sanchez Rodriguez, R Garcia De La Calle, T Perez Vecilla, R Ruiz, F De Guio, M Marina Breysse, M Rienstra, J W BenjaminsAbstract
Introduction
Heart failure with reduced ejection fraction (HFrEF) remains a leading cause of hospitalization and mortality worldwide, frequently diagnosed after symptom onset when structural cardiac damage is already established. Early detection of reduced left ventricular ejection fraction (LVEF) may enable timely intervention. Artificial intelligence (AI) applied to routine electrocardiograms (ECGs) offers a scalable, noninvasive strategy for identifying systolic dysfunction, a precursor to HFrEF.
Purpose
To evaluate the performance of an AI-ECG model in the identification of reduced LVEF (defined as LVEF <40%) using multicentre external validation cohorts.
Methods
A total of 9,740 ECGs with corresponding echocardiograms from multiple European centres were analyzed retrospectively. The single-centre internal validation cohort (n=7,949) underwent structured label refinement to mitigate echocardiographic measurement variability. Model performance was assessed using 100 Monte Carlo Cross-Validation (MCCV) iterations with patient-level splits (70/30). For multicentre external validation evaluation, 1,571 patients with available LVEF measurements were included. Finally, the external validation dataset was also assessed after the addition of a predefined echocardiographic uncertainty interval (40–50% LVEF) for variability-adjusted analysis. Performance was assessed using mean values across MCCV iterations for the area under the receiver operating characteristic curve (AUC), sensitivity and specificity.
Results
In internal validation of reduced LVEF (<40%), the AI model achieved an AUC of 0.977 (95% CI 0.969–0.984), with 91.47% sensitivity (95% CI 88.05–95.10%) and 93.87% specificity (95% CI 90.56–96.36%). External validation yielded an AUC of 0.792 (95% CI 0.786–0.801), with 90.19% sensitivity (95% CI 87.69–92.38%), and 48.87% specificity (95% CI 42.64–54.54%). After addition of the predefined variability interval, the AUC was 0.858 (95% CI 0.848–0.867), with 94.51% sensitivity (95% CI 92.93–95.85%), and 61.31% specificity (95% CI 54.38–67.53%). Receiver operating characteristic (ROC) curves are presented in Figure 1.
Conclusion
This AI-ECG model demonstrates strong internal discrimination and robust external performance to detect reduced LVEF, and supports the potential role of AI-based ECG analysis as a scalable, low-cost approach for early identification of systolic dysfunction. Prospective implementation studies are warranted to evaluate clinical impact and cost-effectiveness.