Data Source
All data in this study was derived from the 2010–2018 American College of Surgeons (ACS) Trauma Quality Improvement Program (TQIP) [15, 16]. More than 450 ACS- and state-verified level I and II trauma centers across North America contribute to TQIP. It includes all patients from verified centers with at least one severe injury (Abbreviated Injury Scale [AIS] ≥ 3 in at least one body region). Data reliability and quality is maintained through training of data abstractors and inter-rater reliability audits of contributing centers.
Research Ethics Board Approval
This study number 20–247 was approved by the Unity Health Toronto Research Ethics Board (Toronto, Ontario, Canada) in January of 2021. Study procedures were followed in accordance with the ethical standards of the institutional committee on human experimentation and with the Helsinki Declaration of 1975. This study used only de-identified retrospective patient data, and individual participant informed consent was waived by the Unity Health Toronto Research Ethics Board.
Study Eligibility
Adult patients (≥ 16 years) with a diagnosis of acute complete (ASIA A) traumatic cervical SCI due to blunt trauma that were treated at level I or II trauma centers were included based on AIS codes (Supplementary Table S1). The International Classification of Diseases 9th and 10th revision Procedure Classification System (ICD-9-PCS and ICD-10-PCS) codes were used to identify procedure codes for decompression and fusion (see Supplementary Table S2). Patients with missing data on whether they underwent spinal surgery were excluded. In addition, patients with missing data on in-hospital mortality were also excluded as this was our primary outcome of interest. Finally, patients with any AIS body score of 6 were also excluded, as these are non-survivable injuries [17].
Patient, Injury, Treatment, And Hospital Characteristics
Several patient and hospital covariates were selected from the TQIP database according to their clinical relevance. Patient demographic data included age, mFI-5, sex, ethnicity, and insurance type. For our analysis we categorized age as 16–60 years, 60–75 years, and > 75 years. The mFI-5 is a frailty index that has been used in trauma and is scored with one point given based on the presence of each of the following: congestive heart failure, diabetes, hypertension, congestive heart failure, chronic obstructive pulmonary disease, and dependent functional status [18]. For our analysis we dichotomized patients into categories of low frailty (mFI < 2) and high frailty (mFI ≥ 2). This type of dichotomy in the mFI-5 has been found to be relevant in prior studies [5]. Sex was dichotomized into male and female, and race was grouped into categories of African American, Caucasian, and other. Insurance was categorized as private, public, and other. Data on the characteristics of the injury were also collected. This included mechanism of injury, presenting Glasgow Coma Scale (GCS), presence of shock (defined as emergency department blood pressure < 90 mmHg), and year of injury. The patient’s GCS was categorized as GCS15, GCS13-14, GCS 9–12, and GCS 3–8, consistent with categories corresponding to severity of traumatic brain injury [19]. Mechanisms of injury were categorized as motor vehicle traumas, falls, and other. The primary treatment characteristic extracted from TQIP was whether the patient underwent a spinal operation. We used ICD-9 and − 10-PCS codes as described above to identify which patients underwent a spinal operation. Surgery was therefore classified as a binary variable. Hospital characteristics including the ACS verification level, teaching status, and hospital size were also extracted from TQIP. Hospital teaching status was categorized as university hospital, community hospital, and non-teaching hospital. Hospital size was categorized as < 200 beds, 200–400 beds, and > 400 beds.
Outcomes
The primary outcome was mortality, which was defined as the presence of an in-hospital mortality during the trauma admission. We computed counts and proportions of mortality across our various age and frailty categories. Using R version 4.2.1 (R Foundation for Statistical Computing, Vienna, Austria), we computed a heat map of the mortality counts using the lattice package [20].
Statistical Analyses
All statistical analyses were performed using R version 4.2.1 with an a priori specified significance level of P = 0.05 (two-tailed). Descriptive statistics were reported as mean and standard deviation (SD) for continuous variables and count and percentage for categorical variables.
Predictive Model And Effect Modification
Logistic regression was used to formulate a predictive model of mortality. Patient age, mFI-5, sex, ethnicity, insurance type, mechanism of injury, presenting GCS, presence of shock, whether they underwent surgery, hospital ACS verification level, teaching status, hospital size, and year of injury were used as covariates for adjustment. To test effect modification of mFI-5 on age, we followed the method described by Wilson et al. and Baron & Kenny [21, 22]. We categorized variables as outlined above, included a final interaction term of age and mFI-5 in the regression model, and considered a moderating effect to exist if the interaction term explained a statistically significant amount of the variance of the outcome variable. To visualize the results of our regression model and the effect modification we used visreg package within R version 4.2.1 (R Foundation for Statistical Computing, Vienna, Austria) [23].
Predictive Model Analysis
Three separate logistic regression models were compared using receiver operating characteristic (ROC) analysis to compare the predictive power of age and mFI-5. We composed a base model of patient, injury, treatment, and hospital covariates described above, and included either age, mFI-5, or age with mFI-5 as covariates in the model. We then computed true and false positive rates of predicted mortality from the regression model to generate an ROC curve. The area under the ROC curve (AUROC) was used to compare a model of age, mFI-5, and age with mFI-5. ROC analysis was conducted using Stata version 17 (Stata Corp, College Station, TX, USA) with an a priori specified significance level of P = 0.05 (two-tailed). We completed pairwise comparisons of the AUROC of the different regression models using a χ2-statistic.
Sensitivity Analyses
We performed sensitivity analyses by computing ROC curves in two patient subgroups. Subgroup 1 was formed by restricting the cohort to patients with age above 60 years. Subgroup 2 was formed by restricting the cohort to patients with age above 60 years, and those who underwent surgery.