The Comprehensive Machine Learning Analytics for Heart Failure

Area of Science:

• Cardiology
• Biostatistics
• Machine Learning

Background:

• Heart failure (HF) is a growing global health concern with increasing prevalence and incidence.
• Early HF detection is crucial for improved patient outcomes, yet diagnosis is challenging due to non-specific symptoms.
• African Americans exhibit a disproportionately higher risk of incident heart failure, necessitating targeted prediction models.

Purpose of the Study:

• To develop and validate a machine learning-based risk prediction model for incident heart failure.
• To specifically address the need for a predictive model tailored to the African American population.
• To optimize model performance by evaluating missing data imputation strategies and predictor inclusion criteria.

Main Methods:

• Implementation of multiple machine learning algorithms: LASSO logistic regression, Support Vector Machine, Random Forest, and Extreme Gradient Boosting (XGBoost).
• Evaluation of predictor inclusion based on varying missing data rates.
• Comparison of different missing data imputation strategies, including non-parametric random forest imputation.

Main Results:

• The Extreme Gradient Boosting (XGBoost) model demonstrated superior predictive performance.
• The optimal XGBoost model achieved an Area Under Curve (AUC) of 0.8409 for heart failure prediction in the Jackson Heart Study cohort.
• Non-parametric random forest imputation and inclusion of variables with <30% missing data yielded the best results.

Conclusions:

• Machine learning, particularly XGBoost, offers a powerful approach for developing accurate heart failure risk prediction models.
• Variations in diabetes medication were identified as a critical, previously underappreciated risk factor for heart failure.
• The developed model provides a valuable tool for early heart failure risk assessment in African Americans, outperforming traditional methods.