Mortality Prediction in Severe Traumatic Brain Injury Using Traditional and Machine Learning Algorithms

Area of Science:

• Neuroscience
• Medical Informatics
• Biostatistics

Background:

• Prognostic prediction for severe traumatic brain injury (sTBI) is critical for clinical management and health policy.
• Existing prediction models for sTBI mortality require validation and potential improvement.

Purpose of the Study:

• To develop and validate advanced prediction models for in-hospital mortality in sTBI patients.
• To compare the performance of machine learning (ML) algorithms against traditional logistic regression (LR) and LASSO regression.

Main Methods:

• Development and validation of LR, LASSO, Support Vector Machines (SVM), and XGBoost models using 54 candidate predictors.
• Internal validation on 2804 sTBI patients from the CENTER-TBI China Registry.
• External validation on 1113 sTBI patients from the CENTER-TBI European Registry.

Main Results:

• XGBoost demonstrated superior discrimination (C-statistic) for mortality prediction compared to LR and LASSO regression.
• The developed XGBoost model outperformed existing models like IMPACT and CRASH.
• XGBoost and SVM achieved high C-statistics (0.87-0.88) in external validation, even with a reduced set of 26 predictors.

Conclusions:

• Machine learning techniques, especially XGBoost, effectively capture complex predictor information in sTBI patients.
• ML models offer more precise mortality predictions for sTBI compared to traditional LR approaches.
• Key predictors include Glasgow Coma Scale score, age, pupillary light reflex, brain Injury Severity Score, and acute subdural hematoma.