Machine Learning Identification of Patient Phenoclusters in Aortic Regurgitation.

BACKGROUND

Current treatment paradigms assume aortic regurgitation (AR) patients to be a homogenous population, but varied courses of disease progression and outcomes are observed clinically.

OBJECTIVES

The aim of this study was to first use unsupervised machine learning to identify unique patient phenoclusters in AR, and subsequently evaluate their prognostic relevance.

METHODS

Clinical and cardiac magnetic resonance (CMR) characterization of moderate or severe AR patients was performed across 4 U.S.

CENTERS

Data from 2 centers were used for derivation of phenoclusters and validation was performed in the other 2. The outcome was all-cause death. An unsupervised clustering pipeline, Partition Around Medoids, used 23 clinical and CMR variables to derive patient clusters independent of outcomes.

RESULTS

Included were 972 patients with mean age 62 ± 23.2 years, 754 (78%) male, 680 (70%) trileaflet valve, and 330 (34%) underwent valve surgery. Over a median follow-up of 2.58 years (Q1-Q3: 1.03-5.50 years), the overall mortality rate was 12%. Four clusters were derived: 1) a younger predominantly male phenotype with majority of bicuspid aortic valve and high extent of left ventricular (LV) remodeling (1% mortality); 2) older male patients with predominantly tricuspid valves and intermediate outcomes (10% mortality); 3) older predominantly male patients with the highest burden of comorbidities, LV scarring, and dysfunction (22% mortality); and 4) a phenotype of predominantly female patients with high mortality and relatively higher symptoms burden, relatively lower extent of LV remodeling, and rate of aortic valve replacement (20% mortality). The clustering algorithm was independently associated with survival after adjustment for time-dependent aortic valve replacement and traditional risk markers of prognosis in patients with AR (C statistic 0.77 vs 0.75; P = 0.009 in the validation cohort).

CONCLUSIONS

Unique patient phenoclusters of AR are described using a machine learning approach leveraging comprehensive CMR and clinical characterization. This approach may be an opportunity for a precision medicine approach to enhance risk stratification of patients with AR. Female patients with AR pose a unique phenotype with high mortality, which deserves greater attention.