Data balancing improves mortality prediction for emergency department patients

Area of Science:

• Emergency Medicine
• Data Science
• Machine Learning

Background:

• Accurate patient mortality prediction in emergency departments (ED) is vital for healthcare services.
• Current models struggle with accuracy and early identification of high-risk patients.
• Machine learning models are sensitive to data imbalance, with fewer mortality samples than survival samples.

Purpose of the Study:

• To enhance patient mortality prediction accuracy in the ED.
• To evaluate the effectiveness of data balancing methods for imbalanced datasets.
• To identify high-risk patients more effectively for improved emergency care.

Main Methods:

• Analysis of 2,437,341 non-traumatic adult ED visits (2008-2016).
• Evaluation of three data balancing methods: Random Under Sampling (RUS), Synthesized Minority Oversampling Technique (SMOTE), and Random Over Sampling (ROS).
• Utilized Random Forest (RF), AdaBoost (ADA), and XGBoost (XGB) with Logistic Regression (LR) as a meta-learner.
• Feature importance analysis using RF, ADA, ADA-BS, and Information Gain (IG).

Main Results:

• XGBoost with ROS achieved the highest AUROC (91.41% for 168-hour mortality), outperforming previous studies.
• Achieved balanced True Positive Rate (TPR) and True Negative Rate (TNR) (79.88% and 86.73%) for practical application.
• ROS outperformed RUS and SMOTE, yielding better AUROC (93.72%) in 24-hour mortality prediction with XGBoost.
• Balanced dataset improved feature importance, particularly for 'Age' and 'Triage' features.

Conclusions:

• The proposed method, utilizing XGBoost and ROS, significantly improves mortality prediction accuracy, especially in challenging long-term timeframes.
• Balanced TPR and TNR offer more practical clinical utility than AUROC alone.
• The data balancing approach enhances the impact of key features like 'Age' and 'Triage' in prediction models.