Enhancing the Prediction of Cardiac Allograft Vasculopathy Using Intravascular Ultrasound and Machine Learning: A Proof of Concept.

Background: Cardiac allograft vasculopathy (CAV) is the leading cause of late graft dysfunction in heart transplantation. Building on previous unsupervised learning models, we sought to identify CAV clusters using serial maximal intimal thickness and baseline clinical risk factors to predict the development of early CAV.

Methods: This is a single-center retrospective study including adult heart transplantation recipients. A latent class mixed-effects model was used to identify patient clusters with similar trajectories of maximal intimal thickness posttransplant and pretransplant covariates associated with each cluster.

Results: Among 186 heart transplantation recipients, we identified 4 patient phenotypes: very low, low, moderate, and high risk. The 5-year risk (95% CI) of the International Society for Heart and Lung Transplantation-defined CAV in the high, moderate, low, and very low risk groups was 49.1% (35.2%-68.5%), 23.4% (13.3%-41.2%), 5.0% (1.3%-19.6%), and 0%, respectively. Only patients in the moderate to high risk cluster developed the International Society for Heart and Lung Transplantation CAV 2-3 at 5 years (=0.02). Of the 4 groups, the low risk group had significantly younger female recipients, shorter ischemic time, and younger female donors compared with the high risk group.

Conclusions: We identified 4 clusters characterized by distinct maximal intimal thickness trajectories. These clusters were shown to discriminate against the development of angiographic CAV. This approach allows for the personalization of surveillance and CAV-directed treatment before the development of angiographically apparent disease.