In our study, we found that older age, lower GCS, higher d-dimer, coagulopathy, hypotension and completely effaced basal cisterns were independent predictors of 1-year mortality in patients with TBI after DC. Compared to the logistic regression model, the random tree model presented a better performance with respect to accuracy, sensitivity, specificity and AUC on the training data (Whether the cut-off point was 0.5 or the optimal point). So did the random tree model in regard to accuracy, sensitivity and specificity on the testing data, when the cut-off point was 0.5. Although the sensitivity of the random tree model was inferior to that of logistic regression model on the testing data at the optimal cut-off point, the accuracy and specificity of the random tree model achieved better. The AUC of the two models is similar on the testing data. On the whole, this finding suggests that the random tree is a valuable and accurate model to predict 1-year mortality in TBI patients after DC. Additionally, our study chose the time of 1-year mortality based on the survival analysis of TBI patients undergoing DC, which showed that the mortality rate within 1 year after discharge was very high. Since we turned the spotlight on the long-term outcomes of TBI patients, patients who died in the hospital were excluded. The predictors of inpatient death and post-discharge mortality were disparate, as the previous study showed9. Thus, our study on 1-year mortality, which excluded patients who died in the hospital, could show better predictive performance to some extent.
Age and GCS, which were already found as important predictors of TBI, were also confirmed in our study13, 18. Tian et al.7 identified that age was one of the independent risk factors for discharge status after DC, and Tang et al.8 also observed that age was the risk factor for 30-day mortality after DC. Combined with our research, older age is a risk factor for both short-term and long-term outcomes of TBI patients after DC. Older people tend to suffer from many basic diseases, and the rehabilitation of the body was poorer after TBI than young people, as we know. GCS, which was similar to age, was also a powerful predictor for outcomes of TBI after DC8, 19. D-dimer, a degradation product of fibrinogen, reflected the fibrinolysis of the body. Many studies found that higher d-dimer at admission was associated with higher risk of progressive hemorrhagic injury20, 21, while a meta-analysis about prognostic role of d-dimer level on admission in TBI patients found no significant relationship between d-dimer and the risk of poor functional outcome at 3 months22. In our study, higher d-dimer was one of the predictors for 1-year mortality after DC. In our opinion, the prognostic role of d-dimer may be related to the study population and specific outcomes. It is considered that secondary coagulopathy after TBI is an important factor for unfavorable outcome23, 24, and our results also confirmed this finding. TBI-induced coagulopathy is very common, ranging from 7–54%25, 26. Coagulopathy generated by TBI is a systemic manifestation of local injury27. Pro-coagulant vesicles (including tissue factors, cardiolipin, vWF, etc.) from damaged brain tissue are released into systemic circulation28–30, making the balance between coagulation and anticoagulation broken. This distinct pathogenetic pathway arises increasing attention, and how to intervene in this process is crucial for the prognosis of TBI patients. Hypotension was another risk factor for 1-year mortality. The prognostic role of hypotension in TBI was poorly elaborated. Tang et al.8 found that intraoperative hypotension was associated with 30-day mortality in TBI patients after DC. In our study, we recorded the incidence of hypotension throughout the course of the disease. Additionally, noradrenaline was injected to maintain vital signs in patients with hypotension. Our study suggests that hypotension is a crucial predictor in long-term prognosis that cannot be ignored in TBI patients after DC. However, some more specific questions between hypotension and the outcome of TBI need to be addressed. For example, whether the course of hypotension in patients with TBI is associated with the outcomes. And the risk factors underlying hypotension in TBI need to be explored. Completely effaced basal cistern status, which represents severe elevated ICP, was found as an important predictor of outcome in the previous study8, 31. Basal cistern effacement is closely associated with pupillary reactivity midline shift. Thus, it can represent a uniquely useful neuroimaging characteristic to guide intervention in TBI32.
Many studies on TBI have been conducted using modern machine learning algorithm, owing to its good prediction performance. Matsuo et al.13 demonstrated that random forest showed good performance for poor outcome prediction at discharge and ridge regression for in-hospital mortality prediction in TBI, both of which achieved almost the accuracy of 0.9. Based on the feature selection method, age and GCS, in their study, appeared to be the most important predictors for both poor outcome and mortality, which was consistent with our findings. A total of 232 patients with TBI were included and separated into training data and test data, which was comparable with our samples of 230 patients. The prediction of mortality was better than our results, which were 0.886 accuracy and 0.875 AUC on the testing data. The difference in performance is mainly due to the prediction of death at different times, and there is no doubt that long-term mortality is harder to predict than in-hospital mortality. Rughani et al.33 used the artificial neural network to predict in-hospital survival of TBI patients, which achieved the accuracy of 0.878 and the AUC of 0.860. They included 11 variables in the model: age, sex, total GCS score, and individual components of the GCS score at the scene of injury and emergency department, and first systolic blood pressure. Nonetheless, some vital parameters, such as biochemical tests, CT scan characteristics and neurologic worsening conditions, were absent in their model. Although one study predicted 18-month mortality in severe TBI after DC using the IMPACT prognostic model19, whose AUC was 0.77, our random tree prediction model achieved the AUC of 0.998 and 0.830 on the testing and training data, respectively. This suggests that machine learning models perform better in outcome prediction than the traditional logistic regression models. At present, machine learning algorithm has been increasingly used in prognosis of TBI13, 34, 35, and it enables us to optimize the treatment strategy and provide better daily care.