Background/Objective:
Traumatic intracranial hemorrhage (tICH) accounts for significant trauma morbidity and mortality. Several studies have developed prognostic models for tICH outcomes, but previous models face limitations, including poor generalizability and limited accuracy. The objective was to develop a prognostic model and determine predictors of mortality using the largest trauma database in the U.S., applying rigorous analytical methodology with true hold-out-set model validation.
Methods:
We identified 248,536 patients in the National Trauma Data Bank (NTDB) from 2012 to 2016 with a diagnosis code associated with tICH. For each admission, we collected demographic information, systolic blood pressure, blood alcohol level (BAL), Glasgow Coma Score (GCS), Injury Severity Score (ISS), presence of epidural/subdural/subarachnoid/intraparenchymal hemorrhage, comorbidities, complications, trauma center level, and trauma center region. Our final study population was 212,666 patients following exclusion of records with missing data. The dependent variable was patient death. Linear support vector machine (SVM) classification was carried out with recursive feature selection. Model performance was assessed using holdout 10-fold cross-validation.
Results:
Cross-validation demonstrated a mean accuracy of 0.792 (95% CI 0.783–0.799). Accuracy, precision, recall, and AUC were 0.827, 0.309, 0.750, and 0.791, respectively. In the final model, high ISS, advanced age, subdural hemorrhage, and subarachnoid hemorrhage were associated with increased mortality, while high GCS verbal and motor subscores, current smoker, BAL beyond the legal limit, and level 1 trauma center were associated with decreased mortality.
Conclusions:
A linear SVM model was developed for tICH, with nine features selected as predictors of mortality. These findings are applicable to multiple hemorrhage subtypes and may benefit the triage of high risk patients upon admission. While many studies have attempted to create models to predict mortality in TBI, we sought to confirm those predictors using modern modeling approaches, machine learning, and true hold-out test sets, using the largest available TBI database in the U.S. We find that while the predictors we identify are consistent with prior reports, overall prediction accuracy is somewhat lower than prior reports when assessed more rigorously.