Data source
This text was conducted using data from the National Trauma Data Bank (NTDB) on patients hospitalized with traumatic injuries between 2012 and 2014 [10]. Available information included patient demographics, AIS codes and ISS (version 2005), mechanism of injury (based on ICD-9-CM E-codes), GCS, length of stay, ICU admission, total number of days spent on the mechanical ventilator, in-hospital mortality, and encrypted hospital identifiers. The dataset consisted of 1,754,977 patients with one or more AIS_05 codes.
We intended to compare different scoring methods based on the version 2005 of the injured AIS predot code. All patients in this research were supposed to have an injury description of the version 2005 for AIS coding. Patients’ age over 89 years (69,478), or younger than 1 year (35,657) were excluded from our analysis. Patients with nontraumatic diagnoses (e.g., drowning/submersion, poisoning, and suffocation), overexertion, or burns were excluded (121,257), missing cause of injury (13,083), missing or invalid data (data missing on age, gender, length of hospital stay, or outcome) (41,269), Patients who were treated only in the Emergency department and not hospitalization were excluded (166,990). Patients who were dead on arrival to the hospital (18,581) or transferred to another facility (71,855) were also excluded. We excluded patients who sustained a single or multiple injuries and AIS_05 severity code component was 9 (5,282). At least 500 trauma patients per hospital annually were available (119,393 patients were excluded). E-codes were mapped to 1 of the 6 mechanisms of injury: stab wound, violence, blunt injury, fall, motor vehicle crash, and firearm wound. The final dataset included 1,198,885 patients admitted to 487 hospitals. Recruitment details were shown in Fig. 1.
Development overview of IMP-2005
In this article, 66.6% of the total data was applied to assess trauma mortality rate (TMR) and weighted median death probability (WMDP) values of different AIS predot codes. When the true mortality rate of a specific AIS predot code was 0, the TMR value was set according to the trend of population crude mortality of each age group in the United States between 2012 and 2014 [11]. The TMR and WMDP values were calculated similar to IMP and IMP-ICDX [7, 12], and were displayed in the Additional files 1 and 2, respectively.
16.7% of the data (IMP-2005 development dataset) was used to evaluate IMP-2005. We applied probit regression model to calculate coefficient of IMP-2005 (Table 4) and deduced specific formula for the IMP-2005. The remaining 16.7% of the data (internal validation dataset) was not applied for the development of WMDP and IMP-2005 to estimate the statistical performance of IMP-2005 and ISS models.
Customized trauma models
This validation dataset provided us with the ability to test the performance of the ISS and IMP-2005. The method to calculate ISS was based on Baker and colleagues [3]. We also computed the total mortality rate for each of the 44 possible ISS values (Fig. 2). The IMP-2005 was defined by four parts. The first part of this model was to incorporate the five most severe (highest) WMDP values as predictors. In order to improve the fitness, we applied a two-term fractional polynomial analysis [13] (WMDP and WMDP3) to each of the first three of these five predictors respectively. The second was to determine whether the worst and second worst traumas were in the same body region. The third was to synthesize the two highest WMDP values into one variable. The last part was NBR (as NBR and NBR0.382, suggested by fractional polynomial transformation [13]). Meanwhile, both models were then reestimated after adding gender, age, GCS, injury mechanism, and mechanical ventilator to simple trauma models, which only include anatomical trauma information.
Statistical analysis
The statistical performance of the trauma models was assessed with the area under the receiver operating characteristic (ROC) curve, the Hosmer-Lemeshow (HL) statistics, and the Akaike information criterion (AIC). The AIC was a measure of the Kullback-Leibler information number, which quantifies how close a statistical model approaches the true distribution. The reason of comparison was that the best model in a particular dataset was the model with the lowest AIC. A bootstrapping algorithm (1,000 replications) was used to calculate the bias-corrected 95% confidence intervals for the ROC and the HL. A P<0.05 was set as statistically significant. All statistical analyses were performed using STATA/MP version 14.0 for Windows. The article was exempted from the examination of the Institutional Review Board of Hangzhou Normal University, People’s Republic of China.