Machine learning-based dynamic mortality prediction after traumatic brain injury.

Area of Science:

• Intensive Care Medicine
• Machine Learning in Healthcare
• Neurological Prognostication

Background:

• Traumatic brain injury (TBI) poses significant mortality risks in intensive care units (ICUs).
• Accurate and timely prognostication is crucial for managing TBI patients.
• Existing prognostic models may lack real-time adaptability.

Purpose of the Study:

• To develop simple, scalable machine learning algorithms for real-time mortality prediction in adult TBI patients.
• To utilize routinely monitored intensive care unit (ICU) data for prognostication.
• To create dynamic predictive models applicable in diverse healthcare settings.

Main Methods:

• Observational, multicenter study involving adult TBI patients monitored for at least 24 hours.
• Logistic regression modeling using machine learning to predict 30-day mortality.
• Algorithms based on intracranial pressure (ICP), mean arterial pressure (MAP), cerebral perfusion pressure (CPP), and Glasgow Coma Scale (GCS), validated using stratified cross-validation.

Main Results:

• Two algorithms were developed: one using ICP, MAP, and CPP; the other including GCS.
• The ICP-MAP-CPP algorithm showed an increase in predictive accuracy (AUC) from 0.67 on day 1 to 0.81 on day 5.
• The ICP-MAP-CPP-GCS algorithm demonstrated higher predictive accuracy, increasing from 0.72 on day 1 to 0.84 on day 5.
• Misclassification occurred in patients who underwent decompressive craniectomy.

Conclusions:

• Novel, simple, and scalable machine learning algorithms for dynamic prognostication in TBI patients are presented.
• The algorithms accurately discriminate between survivors and non-survivors, achieving high predictive values.
• These open-source algorithms offer potential for further development and application, including in low- and middle-income countries.