An Unsupervised Approach to Derive Right Ventricular Pressure-Volume Loop Phenotypes in Pulmonary Hypertension

Abstract

Although right ventricle (RV) dysfunction drives clinical worsening in pulmonary hypertension (PH), information about RV function has not been well integrated in PH risk assessment. The gold standard for assessing RV function and ventriculo-arterial coupling is the construction of multi-beat pressure-volume (PV) loops. PV loops are technically challenging to acquire and not feasible for routine clinical use. Therefore, we aimed to map standard clinically available measurements to emergent PV loop phenotypes. One hundred and one patients with suspected PH underwent right heart catheterization (RHC) with exercise, multi-beat PV loop measurement, and same-day cardiac magnetic resonance imaging (CMR). We applied unsupervised k-means clustering on 10 PV loop metrics to obtain three patient groups with unique RV functional phenotypes and times to clinical worsening. We integrated RHC and CMR measurements to train a random forest classifier that predicts the PV loop patient group with high discrimination (AUC = 0.93). The most informative variable for PV loop phenotype prediction was exercise mean pulmonary arterial pressure (mPAP). Distinct and clinically meaningful PV loop phenotypes exist that can be predicted using clinically accessible hemodynamic and RV-centric measurements. Exercise mPAP may inform RV pressure-volume relationships.

Keywords: cardiac resonance imaging; right heart catheterization; right ventricular‐pulmonary arterial coupling; unsupervised clustering.

Figure 1

Identification of three patient groups with varying clinical outcomes using PV loop features. (A) PCA dimensionality reduction of patients colored by labels found using unsupervised k‐means clustering. (B) Time to clinical worsening curves for each PV loop patient group. (C) Z‐score scaled distributions of PV loop parameters in each PV loop patient group. * indicates p value < 0.05 for multi‐group statistical testing (either one‐way ANOVA or Kruskal–Wallis). (D) Sankey plot for n = 61 patients, describing how COMPERA risk scores map to emergent PV loop patient groups. (E) Sankey plot for n = 60 patients, describing how Reveal Lite 2.0 risk scores map to emergent PV loop patient groups. “L“ indicates low risk, “I“ indicates intermediate risk, and “H“ indicates high risk.

Figure 2

Random forest model performance on predicting PV loop patient group. (A) Receiver‐operating characteristic curve and (B) the precision–recall curve for the random forest model indicate high predictive power when integrating all four data subsets. “AP“ indicates average precision. (C) Testing trained random forest model on data with randomly shuffled labels indicates an accuracy of ~35%.

Figure 3

SHAP analysis of the random forest model indicates top predictive features of the PV loop patient group.