Physiology-informed machine learning for patient-level surgical decision-making in severe traumatic brain injury:...

Area of Science:

• Neurosurgery
• Machine Learning
• Medical Decision Support

Background:

• Prognostic uncertainty complicates surgical decisions in severe traumatic brain injury (TBI).
• Existing models predict overall outcomes but not individual benefit from surgery.
• Physiologic data (vital signs) are underutilized in TBI prognostication and decision support.

Purpose of the Study:

• Develop and validate a machine learning model for estimating individualized benefit from cranial surgery in severe TBI.
• Incorporate vital sign data into the model for enhanced prognostic accuracy.
• Simulate patient-level outcomes to assess the impact of surgical decisions.

Main Methods:

• Retrospective analysis of severe TBI patients (Glasgow Coma Scale ≤ 8) from the TQIP database.
• Utilized causal forest models to estimate individualized treatment effects (ITEs) of surgery.
• Incorporated demographics, injury characteristics, and vital signs; validated using the ROC TBI dataset.

Main Results:

• The model demonstrated strong discrimination for favorable discharge (AUC=0.845) and inpatient mortality (AUC=0.890).
• Implementation of the model increased simulated favorable discharge by 4.1% and reduced inpatient mortality by 13.1% in the TQIP cohort.
• External validation in the ROC dataset showed significant improvements: 10.6% increase in favorable discharge and 43.4% decrease in inpatient mortality.

Conclusions:

• A causal forest-based decision model effectively estimates individualized surgical benefit in severe TBI.
• Incorporating physiologic data enhances personalized, data-driven neurosurgical decision support.
• The model shows potential for improving patient outcomes in severe TBI management.