AI-Aided Triage for GSWH: Validating an Interpretable HCT-Based Mortality Model

Area of Science:

• Neurosurgery
• Radiology
• Artificial Intelligence

Background:

• Civilian gunshot wounds to the head (GSWH) have high mortality rates.
• Current triage lacks standardized, imaging-based tools for GSWH.
• Initial head computerized tomography (HCT) scans are standard but underutilized for prognostication.

Purpose of the Study:

• To develop and evaluate an interpretable, attention-based multiple-instance learning (MIL) model.
• To predict in-hospital mortality using initial HCT scans in GSWH patients.
• To provide objective, reproducible risk estimates and case-level explanations for penetrating brain injury.

Main Methods:

• Retrospective cohort study of 222 adult GSWH patients.
• Development and validation of an attention-based MIL model using stratified random splits.
• Evaluation of model performance on an independent test set (n=54) with 100-fold cross-validation on the development set.

Main Results:

• The MIL model achieved high discrimination for mortality (AUC: 0.92) on the test set.
• Sensitivity was 0.88 and specificity was 0.87 at the optimal operating point.
• Attention maps highlighted critical neuroanatomical regions (brainstem, deep midline, ventricles) correlating with mortality.

Conclusions:

• An interpretable, HCT-based MIL model can accurately predict mortality in GSWH.
• The model provides objective, reproducible risk stratification and transparent explanations.
• This tool supports early prognostication and an imaging-first triage strategy for penetrating brain injury.