Rapid Fire Abstracts
Alper O. Turgut, MD
Resident
University of Pittsburgh Medical Center (UPMC)
Alper O. Turgut, MD
Resident
University of Pittsburgh Medical Center (UPMC)
Jonathan Pan, MD, MBA
Cardiovascular Imaging Fellow
University of Virginia Medical center
Shuo Wang, MD, PhD
Research Associate
University of Virginia Health System
Linda Lee, MD
Assistant Professor, Feinberg School of Medicine
Northwestern Medicine
Connor Wolff
Research Assistant
University of Virginia
Arslan Zahid, MD, MSc
Resident Physician
Emory
Seban Liu, DO
Cardiologist
Riverside Medical Center-Kankakee
Maria Poonawalla, MD
Resident Physician
University of Chicago Medicine
Shruti Hedge, MD
Assistant Professor
Southern Illinois University School of Medicine
Yu Wang, PhD
Graduate Student
University of Virginia
Deyu Sun, PhD
Graduate Student
University of California, Los Angeles
Annie Tsay, MD
Resident
Cambridge Health Alliance
Patrick Norton, MD
Associate Professor
University of Virginia
Roberto M. Lang, MD
Professor of Medicine, Division of Cardiology
University of Chicago Medicine
Christopher M. Kramer, MD
Chief, Cardiovascular Division
University of Virginia Health
University of Virginia
Amit R. Patel, MD
Professor of Medicine
Division of Cardiology, University of Virginia Health System, Charlottesville, Virginia, USA.
Diastolic dysfunction (DD) is an important cause of symptoms and is associated with poor prognosis. However, diastolic function is not routinely assessed during cardiac magnetic resonance (CMR) examinations [1]. There are no clear CMR-based diagnostic criteria for DD, and it is also unknown whether DD is associated with poor prognosis independently of late gadolinium enhancement (LGE). In this study, we used unsupervised machine learning to identify clinical subgroups derived from CMR measurements associated with DD and determined their prognostic value.
Methods:
Patients referred for CMR who had an LVEF > 50% and no LGE were retrospectively enrolled. MACE was defined as composite of heart failure admission, ventricular arrhythmia admission, left ventricular assist device implantation, heart transplant, and cardiac death. CMR images were used to generate left ventricular time-volume curves to determine the early peak filling rate (E-PFR), time to E-PFR indexed to total diastolic time, and time to fill 80% of the LV (Diastolic Volume Recovery or DVR) indexed to total diastolic time (Figure 1). K-means clustering was used with silhouette score evaluation to find the optimal number of clusters. Cox-Proportional Hazards using incidence of MACE and time-to-MACE as labels was assessed. Analysis was performed on Python version 3.10 (Python Software Foundation) using scikit-learn library.
Results:
A total of 324 patients were included with a median follow-up of 5.6 years and total of 30 events. The optimal number of clusters utilizing k-means was 4 clusters. The clinical and CMR characteristics of each cluster identified by k-means analysis can be seen in Table 1. Normal CMR values included high E-PFRs, low DVR times, and low time to E-PFR [2]. As can be seen in the table, cluster 4 had the highest time to E-PFR, highest DVR time, and lowest IEPFR and PFR ratio, as well as large heart volumes. Cluster 4 also had the highest MACE rate amongst the clusters. Amongst patients that clustered to cluster 4, hazard ratios for parameters of DVR (%), LV ESV, and TTPFR (%) were 2.1, 6.4, and 1.34 respectively, indicating high likelihood of MACE with abnormal values of these parameters. Principle component analysis (PCA) shown in Figure 2 illustrates the distribution of patients in each cluster into 4 distinct quadrants and the associated Kaplan-Meier survival curves.
Conclusion:
With an unsupervised machine learning approach applied to CMR parameters of DD, we were able to identify a cluster of higher risk patients who have preserved LVEF and no LGE. These findings support the clinical value of quantifying diastolic function in patients referred for CMR.
