Machine learning based readmission and mortality prediction in heart failure patients

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Heart failure (HF) poses a significant burden on healthcare systems, necessitating accurate prediction of patient outcomes.
• Predicting in-hospital mortality, 6-month mortality, and 30-day/90-day hospital readmissions is crucial for patient management and resource allocation.

Purpose of the Study:

• To develop and evaluate Machine Learning (ML) models for predicting in-hospital and 6-month mortality, and 30-day and 90-day hospital readmissions in heart failure patients.
• To identify the most effective ML approaches and feature selection methods for these prediction tasks using conventional patient data.

Main Methods:

• A cohort of 737 heart failure patients was analyzed using 34 conventional features.
• Data was split into training (70%) and testing (30%) sets, with normalization applied using the Z-score method.
• Feature selection was performed using Boruta, Recursive Feature Elimination (RFE), and Minimum Redundancy Maximum Relevance (MRMR) methods.
• Eight ML models were trained, hyperparameters optimized via tenfold cross-validation and grid search, and evaluated using AUC, accuracy, specificity, and sensitivity.

Main Results:

• The RFE-LR and Boruta-LR models demonstrated high performance for in-hospital mortality prediction (AUC: 0.91 and 0.90, respectively).
• For 30-day rehospitalization, Boruta-SVM and MRMR-LR models showed the best results (AUC: 0.73 and 0.71, respectively).
• Performance for 3-month rehospitalization and 6-month mortality was less robust, suggesting a need for additional data.

Conclusions:

• Reliable ML models were developed for predicting in-hospital mortality and 30-day rehospitalization using conventional features.
• These models can aid in personalizing treatment, improving decision-making, and optimizing healthcare resource allocation.
• Further research incorporating additional data is recommended to enhance prediction accuracy for longer-term outcomes like 3-month readmission and 6-month mortality.