Left Ventricular Ejection Fraction Reporting Variability and AI-Assisted Reproducibility: A Multicentre Analysis

BACKGROUND Artificial intelligence (AI) -assisted assessment of left ventricular ejection fraction (LVEF) has been Increasingly adopted in clinical practice. This research aimed to assess the reliability and reproducibility of AI - assisted LVEF assessment in a diverse, real-world, multicenter setting. METHODS We conducted a retrospective multicenter study involving 354 CMR examinations. A standardized LVEF reassessment (M-LVEF) was performed using QMass 8.1 and systematically compared with original report-derived LVEF values (R-LVEF) generated by vendor-specific AI tools. For inter-observer reproducibility assessment, three operators independently analyzed 30 randomly selected cases using both fully manual and AI-assisted modes, the latter also performed with QMass 8.1. Operator A performed two measurements in both modes for intra-observer analysis. RESULTS Both in overall and single-center analysis, the consistency between R-LVEF and M-LVEF was good or excellent (ICC≥0.86), but the 95% limits of agreement all exceeded ± 5% . 44.3% of cases exhibited >5% differences between R-LVEF and M-LVEF, and 17.5% showed >10% differences. The AI-assisted method not only significantly reduced LVEF assessment time compared with manual analysis, but also demonstrated superior reproducibility. This was evidenced by higher inter-observer agreement among three operators (ICC: 0.968, 0.974, 0.984 for AI vs. 0.913, 0.924, 0.971 for manual) and greater intra-observer reproducibility, all with significantly lower coefficients of variation (5.6%, 8.8%, 9.3% vs 2.5%, 3.8%, 4.4%; 3.5% vs 1.0%, P < 0.05). CONCLUSIONS There are individual differences in AI-assisted LVEF assessment, but it has significant advantages in efficiency and reproducibility, making AI a worthwhile tool to facilitate CMR quantification.