Rapid Fire Session

Session: Rapid Fire Session 4: Machine Learning and Artificial Intelligence, Water-Fat Imaging (Monitor 10)

Reliability and Reproducibility of AI-assisted Left Ventricular Ejection Fraction Assessment: A Multicenter Real-world Retrospective Study

Friday, February 6, 2026

3:52 PM - 3:59 PM

Location: Galapagos / Aruba / Ilhabela (2nd Lower Level)

Presenting Author(s)

yl

yinghui le

physician
Beijing Friendship Hospital, Capital Medical University
Beijing, Beijing, China (People's Republic)

Primary Author(s)

yl

yinghui le

physician
Beijing Friendship Hospital, Capital Medical University
Beijing, Beijing, China (People's Republic)

Co-Author(s)

jz

jiali zhou

physician
Beijing Friendship Hospital, Capital Medical University, China (People's Republic)
SY

Shuna Yang

physician
Beijing Friendship Hospital, Capital Medical University, China (People's Republic)
JA

Jing An, M.D.

doctor
Siemens Shenzhen Magnetic Resonance Ltd, China (People's Republic)
ZL

Zhiqiang Li

attending physician
Binzhou People's Hospital, China (People's Republic)
YL

Yimin Li

attending physician
Zhangzhou Affiliated Hospital of Fujian Medical University, China (People's Republic)
TZ

Ting Zhang

attending physician
Sanmenxia Central Hospital, China (People's Republic)
JZ

Jianying Zhang

attending physician
Nanchong Hospital of Beijing Anzhen Hospital Affiliated to Capital Medical University, China (People's Republic)
JZ

Jixin Zhang

attending physician
Huludao Second People's Hospital, China (People's Republic)
ZJ

Zhiguo Ju

attending physician
Shanghai University of Medicine & Health Science, China (People's Republic)
GL

Guang Li

n/a
Academy of Medical Engineering and Translational Medicine, Tianjin University, China (People's Republic)
JL

Jinchen Li

n/a
Beijing Shuxi Technology Co., Ltd, China (People's Republic)
XS

Xiantao Song

professor
Beijing Friendship Hospital, Capital Medical University, China (People's Republic)
YH

Yi He

Beijing Anzhen Hospital, Capital Medical University, Beijing China, China (People's Republic)
Debiao Li, PhD

Professor
Cedars Sinai Medical Center
Los Angeles, California, United States

Background: Left ventricular ejection fraction (LVEF), serving as a key biomarker for quantitatively assessing left ventricular systolic function, plays a pivotal role in clinical stratification, therapeutic decision-making, and prognostic evaluation of cardiovascular diseases^1-3. This research aimed to assess the reliability and reproducibility of artificial intelligence (AI) - assisted LVEF assessment in a diverse, real-world, multicenter setting.

Methods: We conducted a retrospective multicenter study involving 354 cardiac magnetic resonance (CMR) examinations from six medical institutions. A standardized LVEF reassessment (M-LVEF) was performed by the experienced physicians using an AI-assisted analysis system (QMass 8.1), with systematic comparison against original LVEF assessments from each center's imaging reports (R-LVEF). Three operators with varying levels of experience independently measured 30 cardiac function cases using both fully manual and AI-assisted modes for inter-observer reproducibility analysis. Operator A performed two measurements in both modes for intra-observer analysis. Agreement and correlation of LVEF were evaluated using Pearson correlation coefficient and Bland-Altman analysis, with calculation of bias and 95% limits of agreement (95% LOA). For reproducibility and repeatability analyses, both intraclass correlation coefficient (ICC) and coefficient of variation (CoV) were calculated to evaluate inter-observer and intra-observer variability.

Results: The cohort comprised 221 males (62%) and 133 females (38%), with a mean age of 55 years ± 15. The median R-LVEF was 51% (IQR, 35%-62%), whereas the median M-LVEF was 53% (IQR, 33%-62%; P = 0.007). Both in overall analysis and single-center analysis, the consistency between R-LVEF and M-LVEF was good or excellent (ICC≥0.86, P＜0.001), but the 95% limits of agreement all exceeded ±5% (overall: ranging from −13.8% to 15.9%), see Figure 1. The proportion of cases with >5% differences varied significantly across centers: 60.0%, 56.1%, 32.0%, 39.5%, 38.8%, and 34.7%. The AI-assisted method significantly reduced LVEF measurement time compared with manual assessment (all p< 0.001). AI-assisted assessment of LVEF showed better inter-observer and intra-observer reproducibility than fully manual assessments, with higher ICC values (0.968, 0.974, 0.984 vs 0.913, 0.924, 0.971; 0.998 vs 0.990) and lower coefficient of variation (2.5%, 3.8%, 4.4% vs 5.6%, 8.8%, 9.3%; 3.5% vs 0.7%, P＜0.05), see Figure 2.

Conclusion: There are individual differences in AI-assisted LVEF assessment, but it has significant advantages in reproducibility, making AI a reliable tool to monitor the dynamic changes of cardiac function in patients.