A New Time-Window Prediction Model For Traumatic Hemorrhagic Shock Based on Interpretable Machine Learning.

Area of Science:

• Medical informatics
• Computational biology
• Emergency medicine

Background:

• Traumatic hemorrhagic shock (THS) is a leading cause of mortality and morbidity.
• Early prediction of THS is crucial for timely intervention and improved patient outcomes.
• Existing prediction models may lack accuracy or interpretability.

Purpose of the Study:

• To develop and validate machine learning models for early prediction of THS.
• To evaluate the impact of different feature sets (vital signs, routine blood tests, blood gas analysis) on prediction accuracy.
• To compare the performance of extreme gradient boosting (XGBoost) and logistic regression algorithms.

Main Methods:

• Utilized two large patient datasets: PLA General Hospital Emergency Rescue Database and MIMIC-III.
• Applied extreme gradient boosting (XGBoost) and logistic regression algorithms.
• Developed and tested models using stepped feature sets: vital signs (VS), VS + routine blood (RB), and VS + RB + blood gas analysis (BG) for predicting THS in subsequent time windows (0.5, 1, 2, 3 hours).

Main Results:

• XGBoost models significantly outperformed logistic regression.
• The model using vital signs alone achieved an AUROC of 0.935 at a 0.5-hour prediction window.
• Incorporating laboratory results (RB and BG) improved prediction accuracy, reaching an AUROC of 0.968 at a 1-hour window.
• Models demonstrated good generalization ability in external validation.

Conclusions:

• Machine learning models, particularly XGBoost, can effectively predict THS with high accuracy.
• A combination of vital signs and laboratory data provides the most robust predictions.
• These interpretable models offer a flexible and rolling approach for early THS detection, potentially improving clinical decision-making.