Hospital mortality prediction in traumatic injuries patients: comparing different SMOTE-based machine learning...

Area of Science:

• Medical informatics
• Public health
• Machine learning

Background:

• Trauma is a leading cause of death and disability globally.
• Predicting mortality in intensive care unit (ICU) trauma patients is crucial.
• Imbalanced data poses a significant challenge in developing accurate predictive models.

Purpose of the Study:

• To develop and evaluate Synthetic Minority Over-sampling Technique (SMOTE)-based machine learning (ML) tools for predicting hospital mortality in trauma patients.
• To address the challenge of imbalanced data in trauma patient datasets.
• To enhance the accuracy of mortality prediction models for trauma survivors.

Main Methods:

• A retrospective cohort study of 126 trauma patients admitted to an ICU.
• Application of various SMOTE techniques (SMOTE, Borderline-SMOTE1, Borderline-SMOTE2, SMOTE-NC, SVM-SMOTE) for data preprocessing.
• Utilized Decision Tree, Random Forest, Naive Bayes, Artificial Neural Network, Support Vector Machine, and Extreme Gradient Boosting algorithms for mortality prediction.
• Performance evaluation using metrics such as sensitivity, specificity, accuracy, Area Under the Curve (AUC), F1 score, and G-means.

Main Results:

• SMOTE-based ML algorithms significantly improved prediction performance compared to standard ML algorithms on imbalanced data.
• The mean Area Under the Curve (AUC) for all SMOTE-based models exceeded 91%.
• F1-score and G-means improved from below 70% to over 90% after data balancing with SMOTE techniques.
• Random Forest and Artificial Neural Network (based on SMOTE), and Extreme Gradient Boosting (based on SMOTE-NC) showed the highest performance.

Conclusions:

• SMOTE-based machine learning algorithms offer superior prediction of outcomes in traumatic injuries compared to traditional ML algorithms.
• These enhanced models show potential for assisting ICU physicians in making critical clinical decisions.
• The study highlights the importance of addressing data imbalance for accurate prognostic modeling in trauma care.