2.3.4. Functional Outcome

The Glasgow Outcome Score (GOS) on the day of discharge was estimated based on information from multidisciplinary medical records. The GOS is divided into 5 categories: GOS 1 dead (D), GOS 2 vegetative state (VS), GOS 3 severe disability (SD), GOS 4 moderate disability (MD) or GOS 5 good recovery (GR) [25], and only 2 through 5 were applicable in the present study population. The reasons for reduced GOSs were grouped into (i) TBI, (ii) TBI in combination with extracranial injury and/or comorbidity, and (iii) other.

#### *2.4. Statistics*

Patient characteristics are reported as frequencies (percentages) and means (standard deviations) or medians (interquartile ranges) depending on the distribution. We divided patients into two groups based on the endpoint variable. For comparisons between groups, we used *t*-tests or Mann–Whitney U tests for continuous variables and χ<sup>2</sup> tests or Fisher's exact tests for categorical variables, as appropriate. All tests were two-sided and used the 5% significance level. To examine whether the proportion of patients with moderate– severe TBI who followed the direct pathway increased over the years, logistic regression analysis was used. In the trend analysis, the year variable was treated as an ordinal score. Univariate and multivariate logistic regressions were run to examine independent predictors differentiating between patients discharged to specialized rehabilitation and patients discharged elsewhere. The first model included all patients, and the second model was a subgroup analysis that included patients with moderate–severe TBI. Independent variables were selected based on previous literature and clinical importance. Before conducting the multiple regression analysis, possible multicollinearity of the independent variables

was examined. ICU LOS correlated with ICP sensor (r > 0.68), and GCS at discharge was correlated with GOS (r > 0.66); thus, ICU LOS and GCS were not included in the models. Patients with GOS 5-GR were not eligible for inpatient rehabilitation and hence were not included in the models. Evaluation of the predictive accuracy of the models was assessed by calibration and discrimination. Calibration was evaluated by the Hosmer and Lemeshow goodness-of-fit test, and a statistically nonsignificant result (*p* > 0.05) suggests that the model predicts accurately on average. Discrimination was evaluated by analysis of the area under the receiver operating characteristic curve (AUC ROC). We defined acceptable discriminatory capability as an AUC ROC greater than 0.7 [26]. The results are presented with odds ratios (ORs), 95% confidence intervals (CIs) and *p*-values. Data were analyzed with IBM SPSS Statistics, Version 26.0. Armonk, NY, USA: IBM Corp.

## *2.5. Ethics*

The OUH data protection officer (DPO) approved the Medinsight database (approval number 2016/17569) and this study (approval number 18/20658).
