Abstract 12753: Convolutional Neural Networks Are Non-inferior To Expert Assessment For The Evaluation Of Left Ventricular Systolic Function And Mass In Chronic Heart Failure - Results From The Myovasc Study

Abstract

Introduction: Convolutional neural networks (CNN) can determine markers of cardiac function and structure from echocardiographic imaging datasets. A comparison of measurement results between a CNN and conventional echocardiographic experts with regard to clinical outcome in chronic heart failure (HF) is not available. Methods: Data from the MyoVasc Cohort Study (n=3,289; NCT04064450) on chronic HF were analyzed. Comprehensive clinical phenotyping was performed during an investigation in a dedicated study center. Left ventricular ejection fraction (LVEF) and left ventricular mass (LVM) were measured by a CNN published by Zhang et al. Clinical outcome was assessed by a structured follow-up with subsequent validation and adjudication of endpoints. Results: Data from automatic and human measurements were available in 2,815 subjects with a median follow-up of 5.5 [4.2/6.6] years. The correlation of both measurements was dependent on image quality, ranging between 0.6 and 0.8. In multivariable Cox regression adjusted for age and sex, human and automated LVEF and LVM measures per changes in one standard deviation demonstrated comparable risks for all-cause mortality (LVEF: HRhuman 0.57 [0.52/0.63] vs. HRauto 0.60 [0.55/0.66]; LVM: HRhuman 1.56 [1.42/1.70] vs. HRauto 1.58 [1.43/1.76]) and for worsening HF (LVEF: HRhuman 0.55 [0.51/0.59] vs. HRauto 0.58 [0.53/0.63],; LVM: HRhuman 1.56 [1.45/1.69] vs. HRauto 1.59 [1.47/1.73]). C-indices of both prediction models did not differ significantly (P=0.19). Sensitivity analysis demonstrated no significant differences in model performance for both asymptomatic and symptomatic HF, as well as among HF phenotypes (all P>0.05). Conclusions: CNN-based assessment of LVEF and LVM was non-inferior to conventional expert measurement in predicting all-cause mortality and worsening of HF. This underlines the potential of an CNN-based pipeline for automated evaluation of echocardiography in high-throughput.