Introduction: Current right ventricular (RV) function phenotypes in pulmonary hypertension (PH) rely on few hemodynamic parameters to quantify RV dysfunction disease severity. The aim of the study was to identify novel RV phenotypes using RV hemodynamics and unsupervised machine learning in patients with PH. Methods: Participants were identified from the UArizona PH Registry (n=190). Correlation analyses identified 9 RV variables to include in clustering analysis. Consensus clustering was used to identify an unsupervised clustering method with a large inter-cluster, and small intra-cluster distance (Fig 1A). Significance was tested with Dunn and chi-square tests. Results: K-Medoids with a Pearson distance matrix over >1000 resampling iterations identified 5 clusters (consensus range: 0.8-0.9). World Symposium PH groups were distributed across clusters (Fig 1B). Clusters 1, 2 and 5 exhibit mild RV dysfunction with low diastolic stiffness (Eed) and afterload (Ea, Fig 1C). However, RV-PA coupling (Ees/Ea) was decreased in clusters 1 and 5. Clusters 3 and 4 have moderate to severe RV dysfunction with increased Eed and Ea but cluster 4 had decreased Ees/Ea. For pulmonary hemodynamics, clusters 3 and 4 have increased mean PA pressure (mPAP), increased pulmonary vascular resistance (PVR) and decreased PA compliance (Ca) compared to clusters 1 and 2. Cluster 5 has increased mPAP and high cardiac output resulting in moderately elevated PVR and decreased Ca. Conclusions: Unsupervised clustering identified 5 unique RV phenotypes ranging from mild to severe dysfunction. Further investigations are needed to identify factors modulating intermediate phenotypes to provide insight into RV dysfunction and disease progression.
Read full abstract