Artificial Intelligence for Automated Continuous Quantification of Left Ventricular Ejection Fraction From Standard...

Left ventricular ejection fraction (LVEF) guides diagnosis and treatment across cardiovascular disease, but its measurement from transthoracic echocardiography (TTE) is operator-dependent and subject to observer variability. Artificial intelligence (AI) methods that estimate a continuous LVEF value have proliferated. This systematic review aimed to summarise how closely they agree with reference measurements at the group and individual-patient level. Following PRISMA 2020 guidelines and protocol registration on the International Prospective Register of Systematic Reviews (PROSPERO: CRD420261419064), we searched MEDLINE and Embase (via Ovid), Scopus, and IEEE Xplore from January 2015 to 10 September 2026, with forward and backward citation searching, for peer-reviewed journal articles of AI-derived continuous LVEF from standard adult TTE that reported an error or agreement metric against a reference standard. Risk of bias was assessed with a tailored Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) tool. Results were summarised as unweighted study-level medians without pooling. We included 108 studies. Most were retrospective (86 of 108), and 34 reported an external or independent test cohort as their main result. EchoNet-Dynamic or CAMUS were used in 54 studies. Systematic bias was small (median absolute bias: 1.5 EF percentage points, interquartile range: 0.5 to 3.4; 68 studies). Typical average error was moderate (median mean absolute error: 4.9 points, 62 studies). The median width of the study-specific 95% limits of agreement was 26 points (interquartile range: 22 to 30; 66 studies). Only 10 of 66 studies (15%) had limits within ±10 points, the usual gap between clinical LVEF decision thresholds. The limits were similarly wide when external results were used wherever available (median width: 26.3 points) and somewhat narrower in studies at low risk of bias (23.6 points). AI methods reproduce reference LVEF with small systematic bias, but individual-patient disagreement in most studies exceeds the difference between clinical LVEF categories and the agreement expected between two expert readers. On current evidence, automated LVEF is better suited to assistive use with expert oversight than to stand-alone measurement, pending prospective, externally validated evaluation.