Machine learning models in heart failure with mildly reduced ejection fraction patients.

Area of Science:

• Cardiology
• Medical Informatics
• Biostatistics

Background:

• Heart failure with mildly reduced ejection fraction (HFmrEF) is a distinct clinical phenotype.
• Current guidelines lack robust risk stratification models for HFmrEF mortality and re-hospitalization.
• Accurate prediction is crucial for managing HFmrEF patients and improving outcomes.

Purpose of the Study:

• To develop and validate a novel machine learning (ML) model for predicting mortality and HF re-hospitalization in HFmrEF patients.
• To compare the performance of ML models against traditional risk stratification methods.
• To identify key predictors for HFmrEF patient outcomes.

Main Methods:

• Utilized data from the TOPCAT trial for HFmrEF patients (ejection fraction 45-49%).
• Constructed eight ML models using 72 candidate variables.
• Assessed model discrimination using Harrell concordance index (C-index) and DeLong test; evaluated calibration via bias-corrected estimates.

Main Results:

• LASSO Cox regression demonstrated superior performance for 1- and 6-year mortality prediction (C-indices up to 0.83).
• Random forest (RF) model showed best discrimination for HF re-hospitalization (C-indices up to 0.85).
• Kansas City Cardiomyopathy Questionnaire (KCCQ) subscale scores emerged as the most significant predictor for HFmrEF re-hospitalization.

Conclusions:

• ML-based models significantly outperform traditional approaches in predicting mortality and re-hospitalization for HFmrEF.
• The KCCQ score is a critical factor requiring increased attention in the clinical management of HFmrEF.
• Validated ML models offer a promising tool for personalized risk assessment in HFmrEF.