Dynamic Features Impact on the Quality of Chronic Heart Failure Predictive Modelling

Area of Science:

• Cardiology
• Machine Learning
• Health Informatics

Background:

• Chronic heart failure (CHF) management relies on accurate predictive modeling.
• The impact of temporal patient data (dynamics) on predictive accuracy in CHF is not fully understood.

Purpose of the Study:

• To investigate how patient history and time-varying data (dynamics) influence the quality of predictive models for chronic heart failure.
• To determine the optimal feature sets for various CHF-related prediction tasks.

Main Methods:

• Utilized clinical data from chronic heart failure patients.
• Performed three experiments: CHF episode outcome prediction, CHF stage classification, and heart rate prediction.
• Applied machine learning algorithms (XGBoost, Logistic Regression, Random Forest, Linear Regression, Decision Tree) with static and dynamic features, using forward feature selection.

Main Results:

• The highest prediction quality was achieved when combining both static and dynamic features.
• Dynamic features, representing patient history and evolving states, are crucial for improving predictive accuracy in CHF.
• Feature selection identified optimal combinations for each specific prediction task.

Conclusions:

• Dynamic patient data significantly enhances the performance of machine learning models in chronic heart failure.
• A hybrid approach using both static and dynamic features is recommended for robust CHF predictive modeling.