Automated Left Ventricular Volumetry using Artificial Intelligence in Coronary Calcium Scans (AI-CAC) Predicts Heart Failure Comparably to Cardiac MRI and Outperforms NT-proBNP:

The Multi-Ethnic Study of Atherosclerosis (MESA)

Address for correspondence:

Morteza Naghavi, 2450 Holcombe, Houston, TX, 77021. 650-448-8089. mn@vp.org

AUTHORS (FIRST NAME, LAST NAME):

Morteza Naghavi, Anthony Reeves, Kyle Atlas, Dong Li, Chenyu Zhang, Thomas Atlas, Sion K. Roy, Matthew Budoff, Claudia Henschke, David Yankelevitz, Nathan D. Wong, and Daniel Levy.

a.     HeartLung.AI, Houston, TX

b.     Department of Computer Engineering, Cornell University, Ithaca, NY

c.     The Lundquist Institute, Torrance, CA

d.     Tustin Teleradiology, Tustin, CA

e.     Mount Sinai Hospital, New York, NY

f.      Heart Disease Prevention Program, Division of Cardiology, University of California, Irvine, CA

g.     National Institutes of Health, Bethesda, Maryland

Abstract:

Introduction:

Artificial intelligence-powered coronary artery calcium scan (AI-CAC) provides more actionable information than currently reported. In this study we compared left ventricular (LV) volume measured by AI-CAC versus cardiac magnetic resonance imaging (CMR) and NT-proBNP for predicting heart failure (HF). Additionally, we compared AI-CAC vs. NT-proBNP for detection of left ventricular hypertrophy (LVH) defined by CMR.

Methods:

We used 15-year outcomes data for incident heart failure (HF) from 3078 asymptomatic MESA participants (52.3% women, age 62.2±10.3 years) who underwent both CAC scans and CMR at the baseline examination. Data on CMR semi-manual LV volume, NT-proBNP, and Agatston CAC score were obtained from MESA. Discrimination was assessed using the time-dependent area under the curve (AUC) for incident HF.

Results:

Over 15 years of follow up, 133 cases of HF were diagnosed. The AUC for AI-CAC (0.789) and CMR (0.793) were not significantly different (p=0.67) but were significantly higher than NT-proBNP (0.719) and Agatston score (0.664) (p<.0001) for prediction of incident HF. AI-CAC and CMR significantly improved the continuous Net Reclassification Index of NT-proBNP (0.37) and Agatston score (0.45) for HF prediction (p<0.001 for all). The AUC for AI-CAC vs. NT-proBNP for LVH was 0.871 vs. 0.600 for males and 0.854 vs. 0.600 for females.

Conclusion:

In MESA, AI-CAC automated LV volumetry and CMR semi-automated LV volumetry equally predicted incident HF over 15 years and outperformed NT-proBNP. AI-CAC significantly outperformed NT-proBNP for detection of LVH. Both AI-CAC and CMR significantly improved on NT-proBNP and Agatston CAC score for predicting incident HF.

Keywords: coronary artery calcium, artificial intelligence, cardiac magnetic resonance imaging, left ventricular volume, heart failure.

View Abstract