Dynamic prediction of mortality after traumatic brain injury using a machine learning algorithm.

Area of Science:

• Neuroscience
• Critical Care Medicine
• Biomedical Engineering

Background:

• Optimizing intracranial pressure (ICP) and cerebral perfusion pressure (CPP) is crucial for patients with traumatic brain injury (TBI).
• Dynamic prediction models using time-series data can enhance clinical decision-making in intensive care.

Purpose of the Study:

• To retrain and externally validate a machine learning model for dynamic mortality risk prediction in TBI patients.
• To assess the model's performance using extensive ICP and CPP time-series data.

Main Methods:

• Retrained a machine learning model on 686 TBI patients (62,000 hours of data).
• Externally validated the model in two international cohorts (638 patients, 60,000 hours of data).
• Evaluated model performance using area under the receiver operating characteristic curve (AUC) and precision-recall curves.

Main Results:

• AUC reached 0.79 (Swedish cohort) and 0.73 (American cohort) over time.
• Precision-recall curves improved to 0.57 and 0.64, respectively.
• False positive rates decreased to ≤2.5%.

Conclusions:

• The algorithm provides dynamic, time-evolving mortality predictions for TBI patients in intensive care.
• Model performance improved with increased data availability.
• The tool shows potential as a clinical decision support system for TBI management.