Machine learning vs. conventional statistical models for predicting heart failure readmission and mortality

Area of Science:

• Biomedical Informatics
• Cardiology
• Machine Learning

Background:

• Heart failure (HF) poses a significant challenge in healthcare, with patient readmission and mortality being key indicators of disease severity and treatment effectiveness.
• Accurate prediction of HF readmission and mortality is crucial for timely intervention and improved patient outcomes.
• Machine learning (ML) methods offer potential advancements over conventional statistical models (CSMs) in complex predictive tasks.

Purpose of the Study:

• To systematically review and compare the performance of ML methods against CSMs in predicting readmission and mortality in HF patients.
• To develop and present a framework for evaluating the quality of studies employing ML algorithms for prediction modeling in HF.

Main Methods:

• A systematic literature search was conducted across multiple databases (MEDLINE, EMBASE, etc.) for studies published between January 2000 and July 2020.
• Eligible studies compared ML and CSMs for HF mortality and readmission prediction, with data extracted and quality assessed using a modified CHARMS checklist.
• Twenty articles involving 686,842 patients were included, analyzing various ML techniques (e.g., random forests, neural networks) and CSMs (e.g., logistic regression).

Main Results:

• ML methods demonstrated superior predictive discrimination compared to CSMs in 16 out of 21 readmission comparisons and 7 out of 9 mortality comparisons.
• ML-derived c-indices were consistently higher, particularly in studies using random survival forests for mortality prediction.
• A significant limitation was the lack of external validation in most ML studies; however, the single externally validated study showed ML superiority (c-indices 0.913 vs. 0.835).

Conclusions:

• ML algorithms generally outperform CSMs in predicting HF readmission and mortality.
• There is a critical need for external validation and adherence to clinical quality standards in ML-based prediction modeling research for HF.
• Implementing robust evaluation methods will enhance the reliability and clinical utility of ML predictions in cardiology.