Machine Learning and Artificial Intelligence

1352078 - A Machine Learning Risk Score Based on Routine CMR evaluation to Identify Patients with Left Ventricular diastolic dysfunction and risk of Heart Failure Hospitalization and All-cause Mortality

Thursday, January 26, 2023

10:50 AM – 11:00 AM

Location: Silver Pearl 2

Co-Author(s)

JC

J. Jane J. Cao, MD, MPH

Catholic System Research Director & Professor of Clinical Medicine
St. Francis Hospital, The Heart Center
Greenvale, United States
LW

Lin Wang, MD

MD
St. Francis Hospital, The Heart Center
Roslyn, New York, United States
JC

Jason Craft, MD

Medical Doctor
St. Francis Hospital, The Heart Center, New York, United States
JW

Jonathan Weber, MPH

MPH
St. Francis Hospital, The Heart Center
Roslyn, New York, United States
MP

Michael Passick, MBA

Mr
St. Francis Hospital, The Heart Center, New York, United States
NN

Nora Ngai, PhD

Mrs
St. Francis Hospital, The Heart Center, United States
JG

James W. Goldfarb, PhD

Physicist
St. Francis Hospital, The Heart Center
Woodside, New York, United States
Omar Khalique, MD

Director, Division of Cardiovascular Imaging
St. Francis Hospital, The Heart Center, New York, United States
EB

Eddy Barasch, MD

Medical Doctor
St. Francis Hospital, The Heart Center, United States

Presenting Author(s)

QZ

Qingtao Zhou, PhD

Research Scientist
CHSLI, United States

Disclosure(s):

Omar Khalique, MD: No relevant disclosure to display

Background:

Left ventricular diastolic dysfunction (LVDD) is not currently assessed during routine clinical CMR examinations. In this study we sought to use a machine-learning (ML) risk score to examine the role of CMR in assessing LVDD and to evaluate the risk of heart failure (HF) and all-cause mortality in association with LVDD.

Methods:

We included 606 patients who underwent clinical CMR and echocardiogram within 7 days. LVDD was characterized by echo parameters following the ASE 2016 guidelines. Eight demographic and sixteen CMR variables were used including LV size, mass, ejection fraction (EF), left atrial volume (LAV_min and LAV_max) and total emptying fraction (LAEF) as well the late gadolinium enhancement (LGE). Three machine learning algorithms – Logistic Regression, Random Forest and XGBoost were utilized to predict LVDD. Random Forest was chosen by comparing model performance based on area under curve (AUC) from receiver operating characteristics analysis. The cohort was classified into high and low probability of LVDD based on the ML risk score determined by the Youden Index.

Results:

The subject average age was 66 ± 16 years and 62% were female. There were 305 subjects diagnosed with LVDD following the echocardiographic criteria. The AUC of the model combining the CMR with demographic variables was 0.895 (CI: 0.845-0.939). The study subjects were divided into high (N=297 (49%)) and low ML risk score (N=309 (51%)) groups using a ML risk score cut-point of 0.4121 defined by Youden index, which yielded a sensitivity of 0.82, and a specificity of 0.80 for identifying LVDD. After a mean follow-up of 4.8 years, 123 subjects developed HF requiring hospitalization and 99 died. In the Kaplan-Meier curves below those with high ML risk scores for LVDD were associated with significantly lower event-free survival than those with low ML score (log-rank p< 0.001 for both comparisons).

Conclusion:

Using a ML risk score, it is feasible to identify patients with LVDD using demographic and functional CMR exam and to provide risk stratification for adverse outcomes.