Traumatic brain injury (TBI) is still one of the leading causes of death and disability worldwide with over 10 million people hospitalized every year (1). It is common to witness the alterations of the coagulative system and disturbed coagulation function in TBI patients. Results from previous studies indicated that two in three patients with severe TBI manifested coagulation system abnormalities upon admission to the emergency department, and then continued to worsen (2, 3). And the overall mortality of TBI-induced coagulopathy (TBI-IC) attains 17–86% (4–6). TBI-IC is characterized by both hypo-coagulopathy with prolonged bleeding or hyper-coagulopathy with an increased prothrombotic tendency, or both (4, 7). Previous study unearthed that coagulopathy following TBI was related to higher mortality and prolonged intensive care unit (ICU) stay (8). In early stage, potential mechanisms include the dysfunction of the coagulation cascade and hyperfibrinolysis, both of which contribute to hemorrhagic progression. Later, a poorly defined prothrombotic stage emerges, partly caused by fibrinolysis shutdown and hyperactive platelets (9–11).

Undoubtedly, it is imperative to promote the early identification of TBI-IC in a timely way. Laboratory assays, including international normalized ratio (INR) and thromboelastogram are widely used to diagnose TBI-IC. Nonetheless, these assays have limited value in predicting coagulopathy before it develops. In recent years, as a field of artificial intelligence, machine learning (ML) is able to learn from data based on computational modeling. Likewise, ML can fit high-order relationships between covariates and outcomes in data-rich environments (12–14).

This study aimed to determine whether ML algorithms using demographic, comorbidities, laboratory examinations and other variables could predict TBI-IC with considerable accuracy and identify factors contributing to the prediction power.

Altered hemostasis and hemorrhagic progression are substantial challenges in the clinical management of TBI. Patients with TBI-IC were at a high risk of death over those with normal coagulation. Notably, studies elucidating the rapid prediction of TBI-IC, are warranted. In this sense, our study developed and validated ML models, providing an accurate predictive tool for coagulopathy in TBI patients. Specifically, seven ML models (NNET, NB, GBM, Ada, RF, BT and XGB) were used to predict TBI-IC using variables frequently used in clinical practice. Concerning the predictive performance, the Ada outperformed the remaining models. Moreover, results from the DCA indicated that the Ada and XGB models had higher net benefits over a range of threshold probability values than other models. It is remarkable that this study combined preoperative characteristic, comorbidities, and laboratory findings other than coagulopathy profile to establish a prediction model.

To help surgeons use the model, a calculator was developed, which provided a user-friendly interface. After entering the variables, the incidence of TBI-IC will be shown. The explanation of the ML model at the individual level was consistent with the aforementioned explanations at the feature level, and gratifyingly, the black-box concern was further mitigated to a certain extent. Notably, these results facilitated correct clinical decisions, and more importantly, timely treatment strategy.

A previous study conducted by Cosgriff et al. (20) developed a simple score to predict traumatic brain injury-induced coagulopathy (TIC) using four binary predictors [systolic blood pressure <70 mm Hg, temperature <34°C, pH <7.1, and Injury Severity Score (ISS) >25]. However, due to the fact that the ISS cannot be obtained at the time of decision making, the application of such a score was limited. To predict TIC more accurately, two scores have been developed by prehospital information (21, 22). Mitraet al.'s score used 5 predictors (entrapment; systolic blood pressure < 100 mm Hg; temperature < 34°C; suspected abdominal or pelvic injury; and chest decompression), whereas Peltan et al.'s score employed 6 predictors (age, injury mechanism, prehospital shock index>= 1, GCS, and need for prehospital tracheal intubation and/or Cardiopulmonary Resuscitation (CPR)) (21, 22). Nevertheless, in new patients, both scores achieved only moderate performance, with sensitivity <30%. Additionally, the Trauma Induced Coagulopathy Clinical Score (TICCS) employed three components, including general severity, blood pressure, and extent of significant injuries to predict TIC (23). A major limitation of above scores was that much of the prognostic potential of available information was lost through limiting the number of predictors and dichotomizing continuous variables. Consequently, a novel predictive model for early-identification of TIC was established (Predictors: heart rate, systolic blood pressure, temperature, hemothorax, Focused Assessment with Sonography for Trauma (FAST) result, unstable pelvic fracture, long bone fracture, GCS, lactate, base deficit, pH, mechanism of injury, energy) (24). However, one point worth noting was that previous study focused on the entire trauma patient, not TBI patients in particular, which added confusion to some extent.

By interpreting the full model, it was found that many clinical variables can contribute to predict the risk of TBI-IC. In this study, coagulopathy profile (INR, PT, APTT) was found to be the most important variable in predicting TBI-IC, followed by SOFA, blood routine test (PLT, RBC, HCT, HGB, RDW), renal function (BUN and CRE), comorbidities (congestive heart failure, myocardial infarction) and so forth. Among the fifteen included variables, the SOFA was an important predictor. SOFA is an indicator to describe multiple organ dysfunction, including respiratory system, nervous system, cardiovascular system, liver, coagulation and kidney (25). Potential mechanisms may include the fact that SOFA scores are more likely to indicate liver failure or cardiovascular failure. Those organ failures have a high tendency to bleed, and subsequently leading to coagulopathy (26).

In this study, PLT, RBC, HCT, HGB and RDW were important predictors of TBI-IC. In a prospective observational study conducted by Davis PK et al. (27), PLT dysfunction was an early marker for TBI-IC. Potential mechanism included the blood dilution arised from the use of coagulation factor products (28). Nevertheless, we cannot exclude the likelihood that the blood coagulation system was activated by the continuous bleeding itself (29).

RDW, a parameter of red blood cell volume, measures the variability in size of circulating erythrocytes. Although primarily used to diagnose different types of anemias, the RDW was also associated with various thrombotic disease processes including venous thromboembolism (VTE) (30, 31).

Although the underlying mechanism is unclear, it is speculated that inflammatory factors destroy the vascular endothelial integrity, subsequently changing the glycoprotein and ion channel structure of the erythrocyte membrane (32, 33). Consequently, the deformability of the RBC is reduced, in turn, further enables endothelial damage to increase, causing the release of tissue factors that activate the coagulation pathway and triggers disseminated intravascular coagulation (DIC) (34).

In this study we found that renal function indicators (BUN and CRE) can help to indicate the risk of TBI-IC. Similarly, a ML model developed by Zhao QY et al. also identified renal function, including urine output and CRE to predict sepsis-induced coagulopathy (SIC) (35). It is worth noting that renal dysfunction has been associated with both thrombotic and hemorrhagic complications (36, 37). Potential mechanism included less adenosine diphosphate (ADP) and serotonin storage in PLT of patients with renal dysfunction (38, 39). Taken together, the force of impact at the time of TBI can cause shearing of large and small vessels, and result in subdural, subarachnoid, or intracerebral hemorrhages, or a combination of different types. TBI-associated factors might then alter the intricate balance between bleeding and thrombosis formation, leading to coagulopathy (9). Indeed, the complex interactions between the PLT dysfunction, changes in endogenous procoagulant, anticoagulant factors, endothelial cell activation, hypoperfusion, and inflammation related to TBI-IC remain to be elucidated (9, 40, 41).

The strengths of this study lied in the fact that it applied modern ML approaches to predict TBI-IC. It is worth noting that early and accurate prediction of TBI-IC can provide more time for clinicians to adjust corresponding treatment strategies. For example, this model is applicable if detailed medical history is not available for intubated severe head-injured ICU patient. Furthermore, given the heterogeneity of TBI-IC phenotypes (bleeding/thrombotic tendencies), timely treatment strategy would still require investigation and further testing to determine the type and therefore appropriate treatment. Furthermore, it was based on a real-world data with multicenter and external validation, which heighted the reliability of the performance of ML models. Besides, all the information in this dataset was coded independently of the practitioner, making it a reliable source.

Our study had limitations, consistent with those inherent to many large administrative database studies. First, only TBI-IC adults in ICUs were included, while TBI-IC children and hospitalized TBI-IC cases were not analyzed. Nevertheless, in light of the immaturity of the coagulation system in children, more research is indeed required. Second, derived from the ICU participants, the results of our study cannot be generalized to other population, and we did not obtain information including laboratory testing and interventions before ICU admission, which may cause confounders to some extent. Although our models can screen out patients who are at a high risk of TBI-IC, it is the surgeons who decide the administration of anticoagulant therapy. Usually, the interventions are time sensitive and need to occur early after admission, starting in the emergency department. Third, some new coagulation markers, for example, thrombin-antithrombin-III complex and plasmin-α2-antiplasmin complex, are useful in coagulopathy diagnosis (42, 43). Nevertheless, these indicators were not recorded in the MIMIC-IV and eICU database. This was also the case for viscoelastic coagulation testing [Thrombelastograghy (TEG), Rotational thromboelastometry (ROTEM), ClotPro]. Although these testings can provide detailed coagulopathy diagnosis rapidly and have multiple advantages over the traditional plasma-based coagulation tests (PT, APTT, INR), unfortunately, the above indicators were not included in these two databases. Fourth, we did not obtain the results of cranial Computer Tomography (CT) scans in this study, consequently, the original Corticosteroid Randomization After Significant Head Injury (CRASH)-CT score was not available. Moreover, as an administrative database, there was possibility for misclassification of TBI, to reduce bias caused by imprecise coding, we adopted the extensively used ICD-9, 10 codes. Fifth, as with all potential retrospective studies, there was a potential for unmeasured confounders, causing selection bias. Another major limitation worth noting was the changing nature of the variables in a critically ill patient from time of injury and right throughout the continuum of care to ICU discharge. The nature of the retrospective database did not allow for correction for when measurements were taken in relation to the time of injury. Lastly, although our study deeply explored the coagulopathy of TBI in the ICU settings, other outcomes, such as long-term incidence, are also needed further investigation.

In general, the present study suggested that some important features were potentially related to the TBI-IC. The ML model processed large number of variables and subsequently discriminated TBI patients who would and would not develop coagulopathy, facilitating the implement of timely yet efficient treatments. In the future, further validation regarding its clinical application value will become a necessity.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found below: These data can be found here: https://mimic-iv.mit.edu/; https://eicu-crd.mit.edu/.

The studies involving human participants were reviewed and approved by the Institutional Review Boards of Beth Israel Deaconess Medical Center (Boston, MA) and the Massachusetts Institute of Technology (Cambridge, MA). Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

FY, CP, and LP: Conception and design. YL: Administrative support. JW: Collection and assembly of data. FY and CP: Data analysis and interpretation. All authors: Manuscript writing and final approval.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.