Heart failure with preserved ejection fraction phenogroup classification using machine learning.

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning Applications in Medicine

Background:

• Heart failure with preserved ejection fraction (HFpEF) is a complex syndrome with significant morbidity and mortality.
• Effective treatment strategies for HFpEF require precise patient phenotyping, which remains incompletely understood in Japanese populations.
• Japanese HFpEF patients exhibit lower obesity rates compared to Western cohorts, necessitating population-specific phenotyping approaches.

Purpose of the Study:

• To apply model-based phenomapping using unsupervised machine learning (ML) to identify distinct HFpEF subtypes in Japanese patients.
• To elucidate the clinical characteristics and prognostic implications of identified HFpEF phenogroups.
• To establish a foundation for subtype-dependent treatment strategies in Japanese HFpEF patients.

Main Methods:

• Unsupervised machine learning, specifically a variational Bayesian-Gaussian mixture model (VBGMM), was employed on a derivation cohort of 365 Japanese HFpEF patients.
• Hierarchical clustering was performed, and the VBGMM was validated on an independent cohort of 230 Japanese HFpEF patients.
• Patients were stratified into three phenogroups based on clinical variables, and their 5-year prognosis (all-cause death and HF readmission) was assessed.

Main Results:

• Three distinct phenogroups were identified: Phenogroup 1 (atherosclerosis and chronic kidney disease), Phenogroup 2 (atrial fibrillation), and Phenogroup 3 (younger patients with left ventricular hypertrophy).
• Phenogroup 1 exhibited the poorest prognosis, with a 72.0% incidence of the primary endpoint within 5 years.
• The identified phenogroups were reproducible across derivation and validation cohorts using both VBGMM and hierarchical/supervised clustering methods.

Conclusions:

• Machine learning successfully stratified Japanese HFpEF patients into three clinically relevant phenogroups: atherosclerosis/CKD, atrial fibrillation, and younger/LVH.
• These phenogroups possess distinct clinical characteristics and prognoses, highlighting the potential for personalized medicine in HFpEF.
• The findings underscore the utility of ML-driven phenomapping for advancing understanding and treatment of HFpEF in diverse populations.