Study design and setting
The present study is a multicentre retrospective observational study. It involved the Emergency Departments (ED) of four Italian centers: the Hospital Civile Maggiore of Verona (Italy 100,000 visits per year), the University Hospital of Verona (Italy 50,000 visits per year), the University Hospital of Pisa (Italy 90,000 visits per year) and the General Hospital of Merano (Italy 70,000 visits per year). The study was conducted with the approval of the local ethics committees (Ethics Committee for Clinical Trials, Verona, Italy, approval number 889CESC; Ethics Committee for Clinical Trials, Bolzano, Italy, approval number 75-2019; Ethics Committee for Clinical Trials, Pisa, Italy 11924_CIPRIANO) and was conducted according to the ethical principles for medical research involving human subjects of the Declaration of Helsinki.
Patients
All patients in OAT who required an evaluation in the ED for an MTBI between 1 January 2016 and 31 December 2019, were considered. MTBI was considered as any closed trauma of the cranio-facial district associated with a Glasgow Coma Scale (GCS) of 14-15 at presentation and regardless of loss of consciousness immediately following the trauma [6, 13, 14]. Exclusion criteria were: having access to the ED more than 48 hours after trauma, ineffective OAT, being defined as inadequate intake of Vitamin K Antagonist (VKA) for more than 1 week before the trauma, or having direct oral anticoagulants (DOACs) intake no less than 24 hours before the trauma, having inadequate anticoagulation with VKA, which was defined as International Normalized Ratio (INR) < 1.5.
The records of patients treated with OAT and MTBI were identified according to the following procedure. All patients who underwent cranial CT in the ED during the study period were extracted from the respective computer databases, using dedicated management software (FirstSTATA for Verona and Pisa and QlikView for Merano). The selection of OAT patients only, the congruence with the definition of MTBI, the exclusion criteria, and the recording of baseline and study characteristics were performed with a manual chart review by a group of emergency physicians with more than 5 years of experience.
Clinical management of patients with mild traumatic brain injury
Since 2014, the hospitals under study follow a management protocol for patients with MTBI, based on national guidelines [15]. For patients admitted to the ED for an MTBI in OAT, a head CT is performed on admission and an observation period of not less than 24-h is recommended with the possibility of performing a second head CT before discharge. The protocol also included the collection of pre- and post-traumatic risk factors. Pre-trauma factors were: age ≥ 65 years, presence of antiplatelet therapy, alcohol or drug intoxication on ED arrival, dementia or major psychiatric problems, history of epilepsy, and previous history of neurosurgery. Post-traumatic factors were: major trauma dynamics (defined as a high-speed road traffic accident either as a pedestrian/cyclist or vehicle occupant), fall from a height of more than 3 m, high-speed object accident, post-traumatic transitory loss of consciousness (TLOC), post-traumatic amnesia (any type of post-traumatic amnesia has been considered), presence of post-traumatic headache, presence of signs of trauma above the clavicles, clinical signs of skull base fracture, at least one episode of vomiting, post-traumatic seizure, post-traumatic neurological deficit and a GCS < 15 [6, 15]. The presence or absence of these factors during the patient evaluation was recorded in the ED’s medical chart.
Outcomes
Finding of post-traumatic ICH in head CT scans performed on arrival in the ED (immediate), or in head CT scans performed after 24 h of clinical observation (delayed) was the primary endpoint of the study. CT scan positivity was considered as the presence of subdural, epidural, or parenchymal haematoma, subarachnoid haemorrhage or cerebral contusion [13, 14, 16]. Finally, important patient outcomes were defined as the need for neurosurgical intervention (craniotomy, craniectomy, placement of a hole or subdural drainage) or death from post-traumatic ICH within 30 days of trauma [8, 9, 13, 14, 16]. Patient follow-up was reconstructed by evaluating the medical records available in the computer databases of the EDs in the study, and mortality was confirmed through the registry office.
Decision tree
Decision tree analyses are powerful data mining techniques that are used to classify a set of data and obtain predictions about a dependent output variable. They produce classification and prediction models that help in many decision-making processes. The algorithm underlying the decision trees involves the hierarchical division of the population data into homogeneous subsets according to precise splitting rules. The breakdown of the data, obtained through a set of independent input variables, allows predictions to be made about the target-dependent variable. The decision tree analysis produces a hierarchical diagram consisting of a set of elements called nodes. The node from which subsequent nodes branch is called the root and is composed of the most significant independent variable. Subsequent levels are composed of parent nodes that identify levels and are followed by other nodes at lower levels. Terminal nodes, those that are not further subdivided into other nodes, are also called leaf nodes and can identify subgroups of patients sharing the same risk conditions. The diagram provided by the decision trees is easy to interpret, allows an immediate intuition of the effect of an independent variable on the dependent variable, and, unlike previous regression models, allows an easy interpretation of the relationship between all the variables of the model.
Statistical analysis
Categorical variables were described as percentage and number of events in the total while continuous variables were described as mean and standard deviation (SD) or as the median and interquartile range (IRQ) depending on the underlying distribution. Univariate comparisons were performed with Fischer’s Exact test, Chi-square test, Mann-Whitney test, and Kruskas-Wallis test. Variables found to be significant in the previous univariate analysis (p < 0.05) were entered into the multivariate model. Where appropriate, the Tukey transformation was used to re-express the continuous variables entered in the model, using a power transformation, where the assumption of normality was not guaranteed. A binary logistic regression was used for the multivariate model using the stepwise regression method. Odd ratios with 95% confidence intervals were reported. The study constructed a decision tree using the CART technique. The CART model, a machine-learning and data-mining recursive algorithm were used to identify groups of patients with a homogeneous risk of post-traumatic ICH and to study the hierarchical association between clinical and laboratory risk factors.
Statistical analyses were performed with STATA 16.1 statistical software (StataCorp, College Station, TX, USA).