Development and temporal validation of a two-stage ECG-based machine learning model for LVEF screening: a Middle Eastern cohort study

Le résumé fourni par la source

BACKGROUND: Heart failure is a global health burden, yet artificial intelligence (AI) screening tools for left ventricular ejection fraction (LVEF) have largely been validated in Western populations with limited data from the Middle East. We aimed to develop and validate an interpretable-first, two-stage ECG-based machine learning pipeline to screen for reduced LVEF (Stage 1) and stratify its severity (Stage 2) in a large, diverse Middle Eastern cohort. METHODS: We conducted a retrospective study of 37,233 unique patients from a quaternary-care center in the United Arab Emirates (UAE). The cohort was split into a development set (n = 29,108; 2015–2023) and an independent temporal validation set (n = 8,125; 2024–2025). Stage 1 utilized a logistic regression model to screen for any reduced LVEF (< 52% in men, < 54% in women). Stage 2 applied an XGBoost classifier to positive screens to grade severity as mild (LVEF 41–51% in men, 41–53% in women), moderate (30–40%), or severe (< 30%). RESULTS: Reduced LVEF was present in 18.0% of the overall cohort (17.5% of the development set and 18.6% of the validation set. In the validation cohort, Stage 1 achieved an AUC of 0.82 (95% CI 0.81–0.83) with sensitivity 84% (95% CI 82–86%), specificity 80% (95% CI 78–81%), negative predictive value 91% (95% CI 90–93%), and positive predictive value 68% (95% CI 66–70%). Stage 2 attained 72% overall accuracy; recall for mild, moderate, and severe dysfunction was 74%, 71%, and 72%, respectively, mirroring performance in the internal test set. CONCLUSION: This two-stage AI-ECG workflow provides an interpretable-first and robust method for heart failure screening in a non-Western population. With a high NPV and temporal stability, the model effectively rules out disease and triages severity, offering a scalable “ECG-first” strategy to prioritize echocardiography resources.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.