On-Scene Injury Severity Prediction (OSISP) Algorithm for Truck Occupants

Objective: The aim of this study is to develop an on-scene injury severity prediction (OSISP) algorithm for truck occupants using only accident characteristics that are feasible to assess at the scene of the accident. The purpose of developing this algorithm is to use it as a basis for a field triage tool used in traffic accidents involving trucks. In addition, the model can be valuable for recognizing important factors for improving triage protocols used in Sweden and possibly in other countries with similar traffic environments and prehospital procedures. Methods: The scope is adult truck occupants involved in traffic accidents on Swedish public roads registered in the Swedish Traffic Accident Data Acquisition (STRADA) database for calendar years 2003 to 2013. STRADA contains information reported by the police and medical data on injured road users treated at emergency hospitals. Using data from STRADA, 2 OSISP multivariate logistic regression models for deriving the probability of severe injury (defined here as having an Injury Severity Score [ISS] > 15) were implemented for light and heavy trucks; that is, trucks with weight up to 3,500 kg and ⩾ 16,500 kg, respectively. A 10-fold cross-validation procedure was used to estimate the performance of the OSISP algorithm in terms of the area under the receiver operating characteristic curve (AUC). Results: The rate of belt use was low, especially for heavy truck occupants. The OSISP models developed for light and heavy trucks achieved cross-validation AUC of 0.81 and 0.74, respectively. The AUC values obtained when the models were evaluated on all data without cross-validation were 0.87 for both light and heavy trucks. The difference in the AUC values with and without use of cross-validation indicates overfitting of the model, which may be a consequence of relatively small data sets. Belt use stands out as the most valuable predictor in both types of trucks; accident type and age are important predictors for light trucks. Conclusions: The OSISP models achieve good discriminating capability for light truck occupants and a reasonable performance for heavy truck occupants. The prediction accuracy may be increased by acquiring more data. Belt use was the strongest predictor of severe injury for both light and heavy truck occupants. There is a need for behavior-based safety programs and/or other means to encourage truck occupants to always wear a seat belt.


Introduction
Improving the prehospital care process is fundamental for decreasing mortality and mitigating injury for trauma patients (Murad et al. 2012). Minimizing the delay to definitive treatment has been shown to decrease mortality substantially (Haas et al. 2010). A key to achieve this is to take an early and correct decision on optimal treatment and where to transport the patient. Patients with severe injury should be transported to a trauma center, with the expertise to treat major injury (Haas et al. 2010;MacKenzie et al. 2006). The prehospital personnel's most important decision support is the triage protocol. It facilitates identifying patients with severe On-Scene Injury Severity Prediction 191 injury while using health care resources efficiently by recognizing patients not likely to be in need of urgent and specialized care. Schoell et al. (2014) stated that improvement in triage accuracy is the most promising research field to continue reducing fatal and severe injuries for motor vehicle crashes.
Developing and maintaining a highly accurate triage protocol requires detailed and updated knowledge about what types of traffic accidents have the most severe outcome. Statistics about mortality are rather well established, but there are few publications studying how injury severity is linked to accident characteristics. Mechanisms of Injury (MOI) is used in trauma care to refer to the cause/circumstances of injury (Brown et al. 2011). In this article, the broader term accident characteristics is used to refer to circumstances of the accident such as vehicle characteristics, whether the airbag was deployed, posted speed limit at accident location or occupant's sex and age. A retrospective study including data of around a million trauma patients concluded that using physiologic and anatomic criteria alone results in undertriage and strongly supported the use of MOI in triage systems for trauma (Brown et al. 2011). Hence, the importance of MOI lie in their potential to predict occult injuries to reduce undertriage.
In order to gain understanding of how modern triage systems are designed, the Guidelines for Field Triage of Injured Patients-Recommendations of the National Expert Panel on Field Triage (Sasser et al. 2012) used in United States and the Rapid Emergency Trauma and Triage System (RETTS; Widgren and Jourak 2011), which is the most widespread protocol in Sweden, were reviewed. In both systems, the first step is based on physiological criteria, such as pulse and breathing rate, and the second step is based on anatomical criteria of the identified injuries such as types of fractures. If any of the criteria in the first 2 steps are fulfilled, the patient is given the highest priority level (red in RETTS). Next, in step 3, MOI are assessed. In RETTS, if any of the MOI inclusion criteria are met, the patient is given the second highest priority level (orange). This differs from the U.S. protocol where patients who fulfill MOI criteria should be handled as stated in Sasser et al. (2012;p. 6): "Transport to a trauma center, which, depending upon the defined trauma system, need not be the highest level trauma center." Focusing on the MOI that apply for motor vehicle crashes, some criteria are shared between both triage systems, but differences are also found. Both systems include death in the same vehicle and occupant being ejected from the vehicle. Rollover accident was previously a criterion in the U.S. protocol but it has been removed in its 2 last versions. A similar criterion is used in RETTS: "Vehicle rollover, person trapped." Another criterion in RETTS is the deployment of an airbag. This has no direct equivalent criteria in the U.S. system, which recommends measuring intrusion of the occupant compartment to assess the forces having acted upon the occupant(s). Furthermore, the U.S. expert panel recommends utilizing vehicle telemetry data (transmitted from vehicle to dispatch center) consistent with high risk of injury when available (Sasser et al. 2012). Such data were used in the development of the URGENCY algorithm (Augenstein et al. 2003) for identifying severe crashes and a similar algorithm for the OnStar system (Kononen et al. 2011). It was found that seat belt use, direction and location of impact, and delta V (i.e., the total change in velocity) are important predictors of severe injury that can be recorded by telemetry systems integrated in vehicles (incorporating sensors such as accelerometers and gyroscopes to measure impact forces). However, the use of telemetry systems is currently limited in Sweden, Europe, and most countries outside Europe; thus, algorithms that can predict risk of severe injury as a function of accident characteristics that can be quickly assessed at the scene of accident are valuable. Buendia et al. (2015) showed that an on-scene injury severity prediction (OSISP) algorithm using accident characteristics that are feasible to assess at the scene of the accident by ambulance personnel may be a valuable complement to current triage protocols for victims injured in passenger cars. In that paper, logistic regression was used to relate accident characteristics to the Injury Severity Score (ISS). The area under the receiver operator characteristic (ROC) curve (AUC) was 0.83 for predicting ISS > 15. The most valuable predictors were seat belt use and type of accident. Thus, changing airbag deployment by belt use would represent an improvement for assessment of injury severity in RETTS.
It is unclear from the U.S. and Swedish triage protocols whether the criteria for passenger cars are applicable to truck accidents. RETTS uses the general term vehicle for some MOI criteria and the term car for others. The U.S. protocol refers to "high-risk auto crash," indicating that the MOI criteria are not applicable to trucks. In any case, because truck safety is expected to be different from car safety, different MOI criteria may be appropriate for these groups of patients.
The aim of this study is to develop an OSISP algorithm for truck occupants using only accident characteristics that are feasible to assess at the scene of accident. The purpose of developing the model is to use it as a future complement for field triage for traffic accidents involving trucks.
The OSISP model developed in this article can be valuable for recognizing important factors for improving triage protocols used in Sweden and possibly in other countries with similar traffic environments and prehospital procedures. This is the first study of accidents in Sweden that evaluates how accident characteristics influence the risk of sustaining severe injury for truck occupants using a probabilistic model. A further goal is to identify similarities and differences between the influence of various accident characteristics on the risk of sustaining severe injury for occupants of heavy trucks, light trucks and passenger cars.

Data Selection
The scope of this study is adult truck occupants in traffic accidents registered in the Swedish Traffic Accident Data Acquisition (STRADA) database for calendar years 2003 to 2013. A description of STRADA can be found in Appendix A (see online supplement).
In this study, adult truck occupants (age ≥ 18 years) having both reports available in STRADA-that is, the hospital report for the occupant and the police report of the corresponding accident-were included. The time span considered was from 2003, when the police started to systematically report to STRADA, until and including accidents reported in 2013.
After this selection process, the data were divided into light trucks, with total weight up to 3,500 kg, and heavy trucks, with a total weight over 16,500 kg. These divisions follow the Swedish classification and were used in order to address potential dependence of truck occupant protection on the weight of the truck. Medium weight trucks with a total weight from 3,500 to 16,500 kg were not considered because there were too few cases to construct a prediction model with a reasonable discrimination power (111 patients, of whom 5 had ISS > 15).
The final sample sizes were 2,775 light truck occupants injured in 2,608 accidents and 922 heavy truck occupants injured in 903 accidents.

Injury Severity Assessment and Dichotomization
In order to classify occupants as severely injured or not, the ISS (Baker and OŃeill 1976) was used. ISS is based on the classification of the severity of each injury according to the Abbreviated Injury Scale (AIS; Association for the Advancement of Automotive Medicine 2005). The recommended threshold for triage is ISS > 15 (Sasser et al. 2012); that is, a patient is considered severely injured if ISS > 15 and not severely injured if ISS ≤ 15.

Variables Included in the Model
The dependent variable is whether the patient is classified as severely injured (i.e., ISS > 15) or not. All predictors that were included in the model are detailed in Table B1 in Appendix B (see online supplement).

Data Analysis
All statistical calculations were performed with IBM SPSS Version 22. Univariate chi-squared tests of association were used to compute P values under the null hypothesis of no association between the predictor and the dependent variable. Next, multivariate analyses were performed. Logistic regression modeling was used to discriminate between severe and non severe injury and to compute the odds ratio (OR) of severe injury for each predictor. Logistic regression is a maximum-likelihood method that is commonly used in studies of traffic accidents (see, e.g., Augenstein et al. 2003;Harrell 2001;Kononen et al. 2011;Schiff et al. 2008). A 10-fold cross-validation (CV) procedure was used to estimate the performance of the OSISP algorithm on unseen data, in terms of ROC and AUC. Appendix C (see online supplement) provides details about logistic regression, 10-fold CV, ROC, and AUC.

Results
The results will be presented as follows. For each group-that is, light and heavy truck occupants-the influence of belt use and age on the probability of sustaining severe injury are shown. Afterwards, the 2 binary logistic regression models for light and heavy trucks are derived. For each predictor in these models, statistical significance (P values), OR, and 95% Confidence interval (CI) are shown. Evaluating the performance of the model, AUC and corresponding CI with and without the use of CV are presented. Results for the univariate analysis are presented in Appendix D (see online supplement).

Belt Use and Age Versus Severe Injury
Figure 1 demonstrates a decreasing number of subjects with increasing age. Figure 2 shows the rate of belt use, which is higher in light trucks (79%) than in heavy trucks (55%). The relative frequency of severe injury is substantially lower for belted occupants than for unbelted/unknown ( Figure 2). For both truck types, the relative frequency of severe injury is  higher in the group where belt use was not reported (status unknown) than for unbelted occupants. Figure 2 also shows that for belted occupants, the relative frequency of severe injury is lower for heavy trucks than for light trucks. The same relation holds for those cases where belt use was unknown, whereas unbelted occupants have a slightly higher relative frequency of severe injury for light trucks. Figure 3 shows belt use for each age group. Table 1 shows the results of the multivariate models for occupants in light and heavy trucks. The ROC curves are shown in Figure 4, and AUC values are presented in Table 2. The models achieve CV AUC of 0.81 for light trucks and 0.74 for heavy trucks.

Logistic Regression Models
Belt use stands out as the most valuable predictor for both models. In addition, accident type and age are important predictors for light trucks.
To exemplify the CV performance achieved by the models, if a sensitivity of 95% is required, a specificity of 15% would be obtained for light trucks and 17% for heavy trucks (Figure 4). If otherwise a specificity of 50% is required, a sensitivity of 89% would be obtained for light trucks and 84% for heavy trucks (Figure 4).

Belt Use
A higher overall proportion of severely injured was found for heavy truck occupants (4.0%) than for light truck occupants (2.9%). Furthermore, an even lower probability of severe injury (2.0%) was found in Buendia et al. (2015) for passenger cars. This trend was unexpected because heavier vehicles will in general yield lower delta V in a crash. One explanation for these discrepancies is the large difference in compliance of wearing a seat belt. The rate of belt use for the subjects in the data set was 79% for light trucks and only 55% for heavy trucks, which can be compared to 94% for passenger cars (Buendia et al. 2015). Another important factor may be related to conditional probabilities because only those people who had both a police report and a hospital report were included in the study. There were 12,981 and 8,193 light and heavy truck occupants registered by the police, respectively, which gives a ratio of 1.6. However, there were 2,775 and 922 light and heavy truck occupants remaining after matching hospital and police reports, which gives a ratio of 3.0. This suggests that fewer heavy truck occupants compared to light truck occupants visit a hospital following an accident, for unknown reasons. This difference in conditional probability between the police report samples and the final samples may affect the rates of severe injury reported in this study. Figure 2 shows that for a large proportion of patients, belt use status was unknown and that this group had an even higher rate of severe injury than the unbelted. A possible explanation is that most cases classified as unknown might correspond to injured unbelted occupants who avoided reporting being unbelted for liability reasons. Figure 2 shows that for belted occupants heavy trucks have lower frequency of severe injury (0.79%) than light trucks (1.23%), which can be compared to the frequency for passenger cars of 1.5% found in Buendia et al. (2015). The proportion of severe injury among unbelted occupants is slightly higher for heavy trucks than for light trucks. However, if the belt statuses unknown and unbelted were merged, heavy trucks would have a lower proportion of severe injury than light trucks.
These results indicate that it is important to make truck drivers aware of the necessity to correctly wear a seat belt at all times. In addition to running public campaigns to increase awareness, changing the design of the seat belts and devel- oping smart technology such as (visual and/or auditory) reminders, this can be aided by a concept called behavior-based safety. Behavior-based safety approaches with truck drivers (and their fleet organizations) are based on principles of behavioral change aiming at changing drivers' habits. Professional behavioral change programs are based on engaging, motivating, assisting, reinforcing, and sustaining safe behaviors (e.g., seat belt use). Safety program techniques following the behavior-based safety principles have shown to be relatively easy to implement, cost-effective, and highly efficient for reducing occupational injuries and fatalities in a variety of industrial domains in Wege and Trent (2013).

Predictors
In both types of trucks, belt use is the strongest predictor of severe injury. This conclusion is consistent with the results for passenger car occupants in Buendia et al. (2015).
Regarding age, the number of subjects decreases with age in both types of trucks; see Figure 1. There is a clear trend toward increasing relative frequency of severe injury with age.
Increased proportions of severely injured can be observed for light truck occupants aged > 55 years. This result is consistent with the recommended age partition at 55 years by the expert panel on field triage (Sasser et al. 2012). However, for heavy truck occupants the age group 46 to 55 years old showed a relatively high proportion of injuries. This is likely why elderly occupant-that is, age > 55 years-was a strong predictor for light trucks but that this prediction power decreased for heavy trucks, see Table 1.
Accident type was a very strong predictor for light truck occupants, with head-on accidents having the highest OR, which is consistent with the analysis of passenger cars (Buendia et al. 2015). Note that tram/train accidents were so few that they were not considered in this discussion. Although accident type was not a strong predictor for heavy truck occupants, single-vehicle accidents was the most common type and also produced the highest proportion of severe injury. Volvo Trucks (2013) showed that single-vehicle accidents are the most common accident type in Western Europe producing severe injury to heavy truck occupants, representing 50% of the total number of severely injured cases. In the present study, the proportion was much higher at 84%. For heavy truck occupants, head-on accidents have a lower severe injury probability than rear-end accidents. We do not have a definite explanation for this; possibly, there is a data selection bias because occupants in head-on accidents are more prone to seek/receive hospital care because head-on accidents appear to be more severe compared to rear-end accidents and thus a higher proportion of less severe head-on accidents may be registered in the STRADA hospital sample. However, because the data set is relatively small and only included a few severely injured with the mentioned characteristics, we need to interpret this finding with caution. Airbag deployment was unknown in a large number of cases in both types of trucks. Nevertheless, the models show a higher probability of severe injury when the truck was not equipped with an airbag. This result in itself is insufficient to show the efficiency of airbags because of the presence of potential confounding factors; for example, trucks without an airbag are typically older models and presumably provide a smaller degree of protection than newer models.
Sex, location, and posted speed limit were not strong predictors for any type of truck. Although sex may not be a strong predictor, women experienced a lower probability of being severely injured than men. This result differs from the one reported for cars, where male occupants had a lower probability of severe injury Buendia et al. 2015;Kononen et al. 2011). The lower risk for women is only partially explained by belt use because similar rates as those for men were found; that is, 82% of women versus 79% of men were belted for light trucks, and 61% of women versus 55% of men were belted for heavy trucks.
The injury severity coding in STRADA was changed from AIS 1990 to AIS 2005 in 2007 (see Howard and Linder 2014), and this change generally resulted in a decrease in the coded AIS values for certain injuries. A statistically significant reduction of injury severity was found in the univariate test. However, the period of the accident (2003-2006 or 2007-2013) was not a statistically significant predictor in the logistic regression models. Furthermore, according to the OR in the multivariate model, after adjusting for every predictor, the probability of severe injury increased from 2007 for light truck occupants (Table 1). These results differ from those for car occupants described in Buendia et al. (2015). In that study, the decrement of the probability of severe injury of passenger car occupants produced by the AIS coding change was substantial in both the univariate and multivariate analyses.

OSISP Models for Trucks
The OSISP model developed for light trucks achieved a CV AUC of 0.81. To put this result into perspective, Harrell (2001) stated that a model with an AUC > 0.80 has a good discriminating capability. The discriminating capability achieved for heavy trucks was lower, but the model still performs reasonably well with a CV AUC of 0.74. Comparing the performance with the one reported in Kononen et al. (2011), their model achieved an AUC of 0.84 on self-validation after optimism correction based on bootstrap. That model was based on vehicle telemetry data, and the dependent variable was vehicles with at least one severely injured occupant. Furthermore, the sensitivities and specificities for our model are much higher than the ones reported in Rehn et al. (2012) for a trauma triage protocol in a Norwegian center.
AUC values obtained when the complete data set was used for evaluating the classification performance (i.e., for the full model on self-validation) were 0.87 for both types of trucks. The difference in the AUC values with and without CV indicates overfitting of the model. This may be a consequence of the small data sets available for this study; that is, 2,775 and 922 subjects, of whom only 80 and 37 were severely injured for light and heavy trucks, respectively. In particular, overfitting is more severe for heavy trucks because the data set is smaller, which is reflected in a larger discrepancy between evaluation of the full model and CV than for light trucks.
Overfitting may be reduced by removing the less important predictors or by merging predictor levels with similar OR. However, we decided to include all predictors and levels that were present in Buendia et al. (2015) to allow for a fair comparison between light trucks, heavy trucks, and passenger cars. Because new cases are continuously added to the STRADA database, future analyses based on more data may suffer less from overfitting. The models were developed to be used as a basis for a field triage tool. To achieve the highest triage accuracy possible, overfitting will need to be reduced in future analyses.

Limitations of the Study
We consider the main limitation to be the size of the data set. This causes overfitting and may be the reason behind the lack of statistical significance of certain predictors. In this study belt use, elderly occupant, and accident type yielded P < .05 for light truck occupants, and only belt use yielded P < .05 for heavy truck occupants. This could be compared to the results for car occupants in Buendia et al. (2015) in which all predictors considered in this study, except for airbag deployment, yielded P < .05.
Another limitation is that medium weight trucks were not included because very few accidents were registered in the database used. According to the Swedish classification of trucks, heavy trucks that require a special driver's license are those that exceed 3,500 kg. Heavy trucks are divided according to weight up to 16,500 kg, referred to as medium weight trucks in this study, and above 16,500 kg, referred to as heavy trucks in this study. Volvo Trucks (2013) reported similar results in many aspects for accidents involving heavy and medium weight trucks, but they also found several important differences. Therefore, the results obtained in this article for heavy truck occupants do not necessarily hold for occupants of medium weight trucks, and future studies are needed to develop OSISP algorithms for medium weight trucks.
A number of limitations are related to the underlying database, STRADA. Importantly, belt use in STRADA's hospital report is self-reported by the patient and may possibly be influenced by a perceived need to conform with societal expectations or a fear of liability issues. Furthermore, the requirement of having the hospital report available in STRADA means that only those truck occupants who were admitted to a hospital were included; therefore, the probability of severe injury is actually the conditional probability of severe injury given transportation to a hospital, which may produce a data selection bias. This approximates the unconditional probability of severe injury if the probability of transportation to a hospital is close to one, which may be the case for the target population in which the triage tool is intended to be used. Future prospective evaluations of the OSISP algorithm-that is, field test in prehospital settings that include the actual accidents handled by ambulance personnel-are needed to test the accuracy for the relevant population.
Not all hospitals in Sweden report to STRADA; according to Howard and Linder (2014), the number of hospitals in the system has gradually increased from 29 hospitals in 2003 to 68 hospitals in 2012 out of the 80 hospitals in Sweden in total. This means that the geographical distribution of accidents in the sample may have varied over the 11-year period considered in this study. Finally, STRADA does not provide detailed information about some presumably important variables regarding the accident, like rollover and object impacted. In additions, information about the angle of impact and delta V could potentially have improved model accuracy (Kononen et al. 2011).
Vehicle information about the truck such as model year and weight were not considered due to the perceived difficulty of specifying these variables at the scene of accident. In future OSISP algorithms, vehicle telemetry data could provide such information.