Predicting outcomes after trauma: Prognostic model development based on admission features through machine learning.
Kuo-Chang Lee1, Tzu-Chieh Lin2, Hsiu-Fen Chiang2
1Emergency Department, Chi-Mei Medical Center, Tainan, Taiwan.
Medicine
|December 10, 2021
Summary
A new machine learning model accurately predicts mortality in severe trauma patients using early admission data. This tool aids emergency physicians in critical decision-making for trauma care.
Area of Science:
- Medical Informatics
- Trauma Surgery
- Machine Learning in Medicine
Background:
- Emergency departments (EDs) face challenges in managing severe trauma patients due to the need for rapid, accurate prognostic predictions.
- Existing prognostic tools may not fully leverage early clinical data available within the initial hours of ED presentation.
- Developing an AI-driven model can enhance clinical decision-making for severe trauma outcomes.
Purpose of the Study:
- To develop and validate an early prognostic model for predicting 7-day mortality in severe trauma patients.
- To utilize machine learning, specifically the Extreme Gradient Boosting (XGBoost) algorithm, for enhanced predictive accuracy.
- To identify key admission features and initial ED interventions that are most predictive of patient outcomes.
Main Methods:
- A retrospective analysis of 2232 severe trauma patients (Injury Severity Score >15, age ≥16) from a 4-year database.
- Inclusion of patient data available within the first 2 hours of ED arrival, including Glasgow Coma Scale (GCS), vital signs, and initial interventions.
- Development of an XGBoost model to predict mortality within 7 days of admission.
Main Results:
- The XGBoost model achieved high predictive accuracy (94.0%) and sensitivity (98.0%) for 7-day mortality.
- The model demonstrated a high positive predictive value (PPV) of 95.4%, indicating strong reliability in identifying patients at high risk.
- Key predictors included GCS score, vital signs, prehospital cardiac arrest, abbreviated injury scales (AIS), and specific ED interventions.
Conclusions:
- A machine learning-based prognostic model using early ED data can accurately predict mortality in severe trauma patients.
- The developed model offers a valuable tool for emergency physicians, improving critical decision-making and resource allocation in trauma care.
- High accuracy, sensitivity, and PPV suggest the model's potential for real-world clinical application in trauma management.
