Machine Learning Identification of Patient Phenoclusters in Aortic Regurgitation.

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Aortic regurgitation (AR) patient populations are often treated as homogenous, despite observed variations in disease progression and outcomes.
• Current treatment paradigms may not adequately address the heterogeneity within AR patient cohorts.

Purpose of the Study:

• To utilize unsupervised machine learning to identify distinct patient phenoclusters within the AR population.
• To evaluate the prognostic relevance of these identified phenoclusters.

Main Methods:

• Employed an unsupervised clustering pipeline (Partition Around Medoids) using 23 clinical and cardiac magnetic resonance (CMR) variables.
• Derived patient clusters independently of outcomes using data from 972 patients across 4 U.S. centers, with validation performed.
• All-cause death was the primary outcome measure.

Main Results:

• Four distinct AR phenoclusters were identified with varying mortality rates (1% to 22%).
• Phenocluster 1: Younger males, bicuspid valves, high LV remodeling (1% mortality).
• Phenocluster 2: Older males, tricuspid valves, intermediate outcomes (10% mortality).
• Phenocluster 3: Older males, high comorbidities, LV dysfunction/scarring (22% mortality).
• Phenocluster 4: Predominantly females, high mortality, higher symptoms burden (20% mortality).
• The clustering algorithm independently predicted survival (C-statistic 0.77 vs 0.75, P=0.009 in validation).

Conclusions:

• Machine learning successfully identified unique AR phenoclusters using comprehensive CMR and clinical data.
• This phenotyping approach offers potential for precision medicine and enhanced risk stratification in AR patients.
• A distinct, high-mortality female AR phenotype warrants further investigation and focused attention.