Multiparametric Cardiac MRI for Detection and Grading of Cardiac Allograft Vasculopathy in Heart Transplant Recipients: A Machine Learning Approach (RF_TH_153)

Presenting Author(s)

• 

  Bruno B. Lima, MD, PhD

  Assistant Professor of Medicine
  Vanderbilt University Medical Center

Primary Author(s)

• 

  Bruno B. Lima, MD, PhD

  Assistant Professor of Medicine
  Vanderbilt University Medical Center

Co-Author(s)

• LL

  Luke Laws, MD

  Clinical Fellow
  Vanderbilt University Medical Center
• QW

  Quinn S. Wells, MD

  Associate Professor
  Vanderbilt Univ. Med Ctr
• JW

  John A. Wells, IV, MD

  Assistant Professor of Clinical Medicine
  Vanderbilt University Medical Center
• KA

  Kaushik Amancherla, MD

  Assistant Professor of Medicine
  Vanderbilt University Medical Center
• 

  Tania Ruiz M., MD

  Assistant Professor of Medicine
  Vanderbilt University Medical Center
• MD

  Marianna Daibes, MD

  Postdoctoral Research Fellow
  Brigham and Women's Hospital
• MN

  Malek Nayfeh, MD

  Internal Medicine Resident PGY-1
  Vanderbilt University Medical Center
• CC

  Chloe Carr, MD

  Resident
  Vanderbilt University Medical Center
• SM

  Satish Mishra, MD

  Clinical Fellow, Advanced Cardiac Imaging
  Vanderbilt University Medical Center
• 

  Jonathan Soslow, MD, MSCI

  Professor
  Vanderbilt University Medical Center
  Vanderbilt
• JC

  Jeff Carr, MD

  Cornelius Vanderbilt Professor
  Vanderbilt University Medical Center
• SH

  Sean G. Hughes, MD

  Associate Professor of Clinical Medicine and Radiology Director, Cardiac MRI and Advanced Cardiac Imaging Fellowship
  Vanderbilt University Medical Center
• KS

  Kelly Schlendorf, MD

  Assistant Professor of Medicine and Director, Heart Failure and Transplant Program
  Vanderbilt University Medical Center

Figure 1. Box plots and ROC curves for coronary flow reserve (CFR), stress myocardial blood flow (MBF), native T1, and native T2.

Figure 2. Decision Tree Visualization with Optimal Thresholds.


Background: CAV (Cardiac Allograft Vasculopathy), a leading cause of late mortality after heart transplantation, is typically diagnosed using ICA (Invasive Coronary Angiography), which has procedural risks, especially for patients with renal dysfunction. PET (Positron Emission Tomography) perfusion imaging provides a sensitive method for predicting CAV through CFR (Coronary Flow Reserve) assessment but is limited by high costs and the need for specialized equipment like an on-site cyclotron or Rubidium-82 generator. This study assesses a non-invasive, multiparametric CMR grading system as an alternative to PET for diagnosing CAV.

Methods: This retrospective analysis included 136 heart transplant recipients who underwent stress CMR at Vanderbilt University Medical Center. Patients were stratified by CAV severity defined using International Society for Heart and Lung Transplantation consensus definitions and determined by ICA performed within 30 days of CMR or between two ICAs without a change in CAV classification. CMR-derived parameters were compared between groups and ROC curve analysis was used to identify abnormal thresholds for stress myocardial blood flow (MBF), CFR, T1 and T2.  The dataset was split into training (70%) and testing (30%) sets, and a machine learning decision tree classifier with Scikit-Learn from Python was trained using the identified thresholds to categorize patients into CAV 0, CAV 1, and CAV 2/3. The model's performance was evaluated using confusion matrices, accuracy, and AUC scores.

Results: Table 1 presents the baseline characteristics of patients stratified by CAV severity, showing significant differences in native T1 (p = 0.005), T2 (p = 0.014), and the presence of late gadolinium enhancement (LGE) (p < 0.001).  Abnormal thresholds for key metrics were defined as: stress myocardial blood flow (MBF) ≤ 1.95 mL/min/g, CFR ≤ 1.97, native T1 ≥ 1030 ms, and T2 ≥ 51 ms. The decision tree model (Figure 1) achieved an accuracy of 81.6%. The confusion matrix shows a high sensitivity for CAV 2/3, with 20 out of 23 cases correctly classified. AUC scores were 0.86 for CAV 0, 0.84 for CAV 1, and 0.99 for CAV 2/3, with a macro-average AUC of 0.90 (Figure 2). Feature importance analysis revealed that CFR total (40.3%), ischemia (22.5%), and non-ischemic LGE (22.1%) were the most significant predictors of CAV severity.

Conclusion: Our results suggest that multiparametric CMR can non-invasively detect moderate-to-severe CAV with high accuracy and underscore the importance of tissue characterization and measurement of CFR and MBF in CAV assessment. Specifically, CMR may be utilized to screen for heart transplant recipients at high-risk for CAV 2/3 and inform decisions regarding ICA in that group. Multiparametric CMR offers a non-invasive, accurate screening method for CAV that could significantly reduce the frequency of surveillance ICA, and thus the associated morbidity, in patients after transplant. Further research is needed to refine this model and confirm its utility across diverse populations.

Figure 1. Box plots and ROC curves for coronary flow reserve (CFR), stress myocardial blood flow (MBF), native T1, and native T2.

Figure 2. Decision Tree Visualization with Optimal Thresholds.