Characterizing advanced heart failure risk and hemodynamic phenotypes using interpretable machine learning.

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Existing risk models for advanced heart failure with reduced ejection fraction (HFrEF) often lack integration of invasive hemodynamics and robust handling of missing data.
• This study addresses the need for advanced risk stratification tools in HFrEF management.

Purpose of the Study:

• To develop and validate a heart failure (HF) hemodynamic risk and phenotyping score for HFrEF using interpretable Machine Learning (ML).
• To predict patient risk categories (1-5) based on a composite endpoint, utilizing both invasive and comprehensive feature sets.

Main Methods:

• Patients were categorized into 5 risk groups using unsupervised clustering based on a composite endpoint (death, LVAD, transplantation, rehospitalization within 6 months).
• Interpretable ML models were developed to predict these risk categories using invasive hemodynamics or a comprehensive feature set including non-invasive data.
• Models were trained on the ESCAPE trial data and validated across four additional advanced HF cohorts.

Main Results:

• The developed ML models demonstrated high accuracy in predicting patient risk categories across all outcomes.
• Prediction accuracies for risk categories ranged from 0.896 to 0.969 for the invasive hemodynamics set and 0.858 to 0.997 for the all features set.
• The models proved robust to missing data, enhancing their clinical applicability.

Conclusions:

• Novel interpretable ML models accurately predict distinct risk categories in HFrEF, offering a new paradigm beyond binary outcome prediction.
• This approach facilitates nuanced risk stratification, potentially guiding personalized treatment selection for advanced HF patients.
• Further prospective clinical evaluation is recommended to ascertain the utility of this risk stratification method in clinical practice.