Fluid volume status detection model for patients with heart failure based on machine learning methods.

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Fluid volume abnormalities are a primary cause of heart failure exacerbations.
• Current methods for assessing fluid status are often inefficient, slow, and costly.
• Accurate volume status assessment is crucial for effective heart failure management.

Purpose of the Study:

• To develop and validate an early fluid volume detection model for heart failure patients.
• To utilize machine learning for stratifying patient fluid volume status.
• To improve the accuracy and efficiency of fluid status assessment in heart failure.

Main Methods:

• Utilized a dataset of 2056 heart failure patients with 97 medical characteristics.
• Applied Minimum Redundancy Maximum Relevance (mRMR) for feature selection.
• Developed and evaluated four machine learning classification models using ROC AUC, calibration curves, and other metrics.
• Externally validated models with 186 heart failure patients.
• Employed SHapley Additive exPlanations (SHAP) for model interpretability.

Main Results:

• Selected 30 key features for model development.
• Achieved an Area Under the ROC Curve (AUC) ranging from 0.64 to 0.77 across models and datasets.
• External validation demonstrated model performance with AUCs between 0.64 and 0.74.
• SHAP analysis provided global interpretations of logistic regression models.

Conclusions:

• Machine learning models demonstrate efficacy in detecting fluid volume status in heart failure patients.
• These models can serve as valuable tools for assisted diagnosis.
• The findings support the use of machine learning to enable precise, tailored management strategies for heart failure patients.