Predicting Cardiac Allograft Vasculopathy Profiles Using Machine Learning Clustering

Purpose Cardiac Allograft Vasculopathy (CAV) is the leading cause of graft dysfunction in heart transplantation (HT). Identifying CAV profiles is critical in determining the intensity of surveillance required to appropriately prevent and treat this complication. We sought to identify CAV profiles using maximal intimal thickness (MIT) and a latent-class mixed-effect regression model. Methods This analysis includes an adult HT patients from a single centre from 2010 to 2018. Latent class mixed-effect regression models were used to simultaneously identify latent classes (or clusters) of MIT trajectories post-HT and to quantify risk factors associated with the cluster membership. The outcome of interest was MIT from baseline (0-6 months) to 5 years after HT. We considered recipient and donor age, sex, and ischemic time as risk factors. We subsequently characterized the association between the clusters and ISHLT CAV grade using a competing risk model. Results Overall 186 consecutive adult HT recipients, predominantly male (70%), Caucasian (58%), 54 (43-61) years of age were included. In a covariate-adjusted model, 4 CAV profiles were identified (Figure 1A). Class 1 and 2 consisted of patients with stable MIT over time; patients in Class 3 and 4 experienced elevated MIT. Patients in Class 3 had a lower MIT and slower progression than those in Class 4. There were significant between-class differences in recipient age (p=0.002), recipient sex (p= Conclusion This work builds on previous ML models to differentiate trajectories of CAV using MIT. In our model, MIT at baseline and year-1, in addition to baseline demographics, can be used to discriminate the progression of CAV at 5 years. Recipients in CAV Class 3 and 4 would benefit from more intensive surveillance and early CAV directed therapies.