Prediction of mortality in heart failure by machine learning. Comparison with statistical modeling

Area of Science:

• Cardiology
• Biostatistics
• Machine Learning

Background:

• Predicting heart failure prognosis using machine learning (ML) and conventional statistical methods is a complex research area.
• Accurate prognostic models are crucial for guiding clinical decisions in heart failure management.

Purpose of the Study:

• To compare the performance of various ML models against established scores and logistic regression for predicting heart failure mortality.
• To evaluate the predictive capabilities of Random Forest (RF) and Extreme Gradient Boosting (XGBoost) in heart failure prognosis.

Main Methods:

• Five ML approaches (RF, Gradient Boosting, XGBoost, Support Vector Machine, Multilayer Perceptron) were employed.
• Model performance was assessed using discrimination, calibration, and net benefit metrics.
• ML models were compared against the MAGGIC score and a novel logistic regression model (LRM).

Main Results:

• XGBoost and RF demonstrated strong performance, outperforming the MAGGIC score.
• XGBoost achieved the highest discrimination (C-statistic: 0.793), while RF excelled in precision-recall.
• The logistic regression model (LRM) showed comparable or superior performance to the best ML models.

Conclusions:

• RF and XGBoost models are effective in predicting heart failure mortality, surpassing the MAGGIC score.
• Despite their performance, ML models did not provide a significant advantage over a logistic regression model using the same variables.