Developing and validating a machine learning-based model for predicting in-hospital mortality among ICU-admitted
De Su1, Jie Zheng1, Yue-Kai Shao1
1Department of Critical Care Medicine, Affiliated Hospital of Zunyi Medical University, Zunyi, Guizhou, P.R. China.
Digital Health
|April 29, 2025
Summary
Machine learning models, especially ensemble methods, can predict heart failure patient mortality in the ICU. These models offer a valuable tool for clinical decision-making and improving patient outcomes.
Area of Science:
- Cardiology
- Medical Informatics
- Machine Learning
Background:
- Accurate prediction of in-hospital mortality for heart failure patients in the intensive care unit (ICU) is critical for clinical decision-making.
- Current predictive models lack comprehensiveness in assessing prognosis for these patients.
- Machine learning (ML) offers a powerful approach to identify risk factors and predict outcomes using complex clinical data.
Purpose of the Study:
- To develop and evaluate machine learning models for predicting in-hospital mortality risk in intensive care unit (ICU) heart failure patients.
- To compare the performance of various ML algorithms, including ensemble methods, for this prediction task.
Main Methods:
- Utilized the MIMIC-III database to extract demographic, vital signs, laboratory, and comorbidity data for heart failure patients.
- Employed LASSO regression for feature selection and trained models using logistic regression (LR), random forest (RF), gradient boosting (GB), and other ML algorithms.
- Constructed a soft voting ensemble learning model and evaluated performance using accuracy, recall, precision, F1 score, and AUC via cross-validation and an independent test set.
Main Results:
- The soft voting ensemble model achieved the highest performance in five-fold cross-validation, with accuracy and AUC both at 0.86.
- On an independent test set, the random forest (RF) model yielded an accuracy of 0.79 and AUC of 0.79, while the gradient boosting (GB) model achieved 0.77 accuracy and 0.79 AUC.
- Other models like logistic regression (LR), support vector machine (SVM), and k-nearest neighbors (KNN) showed lower accuracy and AUC, highlighting the advantage of ensemble methods.
Conclusions:
- Machine learning models, particularly soft voting ensemble models, show significant potential for predicting in-hospital mortality in ICU heart failure patients.
- The developed ensemble model serves as an effective adjunct tool for clinical decision-making.
- Further model optimization and validation across diverse patient populations are recommended to enhance clinical utility and accuracy.
