Predicting risk for trauma patients using static and dynamic information from the MIMIC III database.

Area of Science:

• Critical Care Medicine
• Health Informatics
• Machine Learning in Healthcare

Background:

• Risk quantification algorithms in Intensive Care Units (ICUs) are crucial for early alerts and efficient resource management.
• Electronic health records enable the development of predictive models for patient risk stratification.

Purpose of the Study:

• To train a gradient boosting classifier to predict high-risk and low-risk trauma patients.
• To evaluate the model's ability to distinguish between patient risk levels using static and dynamic variables.

Main Methods:

• Utilized the MIMIC-III database, extracting 5,400 trauma patient records with static and dynamic variables.
• Trained a gradient boosting classifier using 3-hour moving time windows of dynamic variables (mean, standard deviation, skew).
• Incorporated an admission survival metric from a National Trauma Data Bank (NTDB)-trained model.

Main Results:

• The final model achieved an Area Under the Receiver Operator Characteristic Curve (AUROC) of 92.9% in distinguishing high-risk from low-risk trauma patients.
• Dynamic survival probability plots showed distinct trajectories for patients who died versus those who survived.
• Demonstrated effective reduction of high-dimensional patient data into a single trauma trajectory.

Conclusions:

• Gradient boosting models can accurately predict high-risk trauma patients in the ICU.
• Dynamic survival probability analysis offers a valuable tool for visualizing and understanding patient risk trajectories.
• This approach aids in clinical decision-making and resource allocation for critically ill trauma patients.