Physiology-informed machine learning for patient-level surgical decision-making in severe traumatic brain injury: Multicenter model development with external validation.

BACKGROUND: Surgical decisions for severe traumatic brain injury (TBI) are often made under prognostic uncertainty. Existing prognostic models predict overall outcomes but do not estimate how an individual's outcome might change with the decision to operate. Moreover, physiologic data such as vital signs are often collected but rarely used for such decision support. We aimed to develop and externally validate a machine learning-based decision model including vital sign data to estimate individualized benefit from cranial surgery and simulate patient-level outcomes.

METHODS: Adults with severe TBI (Glasgow Coma Scale ≤ 8) were retrospectively identified in the TQIP database. The primary exposure was cranial surgery, and outcomes were favorable discharge disposition and inpatient mortality. Causal forest models estimated individualized treatment effects (ITEs) of surgery on favorable discharge, incorporating demographics, injury characteristics, and prehospital and emergency department vital signs. Random forest classifiers predicted counterfactual outcomes under surgery versus no surgery. Model discrimination, calibration, and optimism-corrected performance were assessed using nested cross-validation and bootstrap resampling. The optimal surgical decision threshold was identified by maximizing favorable discharge. External validation was performed using the Resuscitation Outcomes Consortium (ROC) TBI dataset.

RESULTS: Among 10,782 TQIP patients with severe TBI (median age, 48 years), 23% underwent cranial surgery. Incorporating vital signs into the causal forest model allowed for the estimation of more clearly-defined individualized treatment effects. Random forest models showed strong discrimination (favorable discharge AUC = 0.845; inpatient mortality AUC = 0.890) and good calibration. Implementing the surgical decision model increased simulated favorable discharge by 4.1% and reduced inpatient mortality by 13.1%. External validation in 182 ROC patients with TBI showed similar improvements (favorable discharge, 10.6% increase; inpatient mortality, 43.4% decrease).

CONCLUSIONS: A causal forest-based decision model incorporating physiologic data enabled individualized surgical benefit estimation and improved simulated outcomes in severe TBI, supporting its potential for personalized, data-driven neurosurgical decision support.