Improving risk prediction in heart failure using machine learning.

Area of Science:

• Cardiology
• Biostatistics
• Machine Learning

Background:

• Accurate mortality prediction is crucial for heart failure (HF) patients.
• Current risk prediction strategies for HF have limited success.
• Existing methods fail to capture complex interactions in large datasets.

Purpose of the Study:

• To develop and validate a machine learning-based mortality risk score for heart failure (HF) patients.
• To identify key variables for predicting mortality risk in HF.
• To compare the performance of the new risk score against existing methods.

Main Methods:

• A boosted decision tree machine learning algorithm was trained on a cohort of 5822 HF patients.
• The model identified eight variables: diastolic blood pressure, creatinine, blood urea nitrogen, hemoglobin, white blood cell count, platelets, albumin, and red blood cell distribution width.
• External validation was performed in two independent HF populations.

Main Results:

• The derived risk score achieved an area under the curve (AUC) of 0.88 in the training cohort.
• The score demonstrated strong predictive performance across the full spectrum of mortality risk.
• External validation yielded AUCs of 0.84 and 0.81, outperforming two existing risk scores.

Conclusions:

• A novel, validated mortality risk score for HF patients was generated using machine learning and accessible variables.
• This machine learning approach offers superior accuracy compared to current risk scores.
• The findings support the application of this machine learning strategy for HF risk assessment and other challenging predictive settings.