Artificial intelligence in ECG-based diagnosis of low left ventricular ejection fraction: a systematic review and meta-analysis.

Abstract

Detecting Left Ventricular Systolic Dysfunction (LVSD) is crucial for counteracting heart failure progression. While Electrocardiograms (ECG) are widely used, their standalone diagnostic accuracy is insufficient. Integrating Artificial Intelligence (AI) with ECG analysis offers a promising approach to enhance precision. A systematic review was conducted to assess AI-enabled ECG for LVSD detection. Of 394 initial studies, 19 qualified for the systematic review, with 17 incorporated into meta-analysis. Study quality was gauged using QUADAS-2. Univariate meta-analysis, Spearman correlation, and bivariate meta-analyses were performed, along with publication bias assessment. The pooled sensitivity and specificity for AI-enabled ECG models were 86.9% and 84.4%, respectively. Studies with an ejection fraction (EF) threshold of 35% had the highest sensitivity, while those with 50% showed lower sensitivity and specificity. A weak positive Spearman correlation was found across all studies (ρ = 0.374, p = 0.066). A strong positive correlation for externally validated studies (ρ = 0.696, p = 0.008), and a weak negative correlation for test-only studies, indicated a threshold effect. Hierarchical summary receiver operating characteristic curve showed diagnostic robustness for studies with a 40% EF threshold; however, it showed a lack of real-world generalizability for test-only studies. AI-enabled ECG models show strong diagnostic potential for severe LVSD but remain limited for mild cases. External validation is essential for robustness and generalizability. Future research should enhance diagnostic accuracy for mild LVSD and address publication bias to optimize AI-based tools.