Death After Discharge: Prognostic Model of 1-year Mortality in Traumatic Brain Injury Patients Undergoing Decompressive Craniectomy

Abstract

OBJECTIVE: Despite advances in decompressive craniectomy (DC) for the treatment of traumatic brain injury (TBI), a high risk of poor long-term prognosis exists in these patients. The aim of this study is to predict 1-year mortality in TBI patients undergoing DC using the logistic regression and random tree models.METHODS: This was a retrospective analysis of TBI patients undergoing DC from January 1, 2015 to April 25, 2019. Patient demographic characteristics, biochemical tests and intraoperative factors were collected. 1-year mortality prognostic models were developed using multivariate logistic regression and random tree algorithms. Overall accuracy, sensitivity, specificity and area under the receiver operating characteristic curves (AUC) were used to evaluate model performance.RESULTS: Of the 230 patients, 70 (30.4%) died within 1 year. Older age (OR, 1.066; 95% CI, 1.045-1.087; P < 0.001), higher Glasgow coma score (GCS) (OR, 0.737; 95% CI, 0.660-0.824; P < 0.001), higher d-dimer (OR, 1.005; 95% CI, 1.001-1.009; P = 0.015), coagulopathy (OR, 2.965; 95% CI, 1.808-4.864; P < 0.001), hypotension (OR, 3.862; 95% CI, 2.176-6.855; P < 0.001) and completely effaced basal cisterns (OR, 3.766; 95% CI, 2.255-6.290; P < 0.001) were independent predictors of 1-year mortality. Random forest demonstrated better performance for 1-year mortality prediction, which achieved overall accuracy of 0.810, sensitivity of 0.833, specificity of 0.800, and AUC of 0.830 on the testing data, compared to logistic regression model.CONCLUSIONS: Random forest model showed relatively good predictive performance for 1-year mortality in TBI patients undergoing DC. Further external test is required to verify our prognostic model.