Development of a concise injury severity prediction model for pediatric patients involved in a motor vehicle collision

Abstract

Objective:

Transporting severely injured pediatric patients to a trauma center has been shown to decrease mortality. A decision support tool to assist emergency medical services (EMS) providers with trauma triage would be both as parsimonious as possible and highly accurate. The objective of this study was to determine the minimum set of predictors required to accurately predict severe injury in pediatric patients.

Methods:

Crash data and patient injuries were obtained from the NASS and CISS databases. A baseline multivariable logistic model was developed to predict severe injury in pediatric patients using the following predictors: age, sex, seat row, restraint use, ejection, entrapment, posted speed limit, any airbag deployment, principal direction of force (PDOF), change in velocity (delta-V), single vs. multiple collisions, and non-rollover vs. rollover. The outcomes of interest were injury severity score (ISS) ≥16 and the Target Injury List (TIL). Accuracy was measured by the cross-validation mean of the receiver operator curve (ROC) area under the curve (AUC). We used Bayesian Model Averaging (BMA) based on all subsets regression to determine the importance of each variable separately for each outcome. The AUC of the highest performing model for each number of variables was compared to the baseline model to assess for a statistically significant difference (p<0.05). A reduced variable set model was derived using this information.

Results:

The baseline models performed well (ISS≥16: AUC 0.91 [95% CI: 0.86–0.95], TIL: AUC 0.90 [95% CI: 0.86–0.94]). Using BMA, the rank of the importance of the predictors was identical for both ISS≥16 and TIL. There was no statistically significant decrease in accuracy until the models were reduced to fewer than five and six variables for predicting ISS≥16 and TIL, respectively. A reduced variable set model developed using the top five variables (delta-V, entrapment, ejection, restraint use, and near-side collision) to predict ISS≥16 had an AUC 0.90 [95% CI: 0.84–0.96]. Among the models that did not include delta-V, the highest AUC was 0.82 [95% CI: 0.77–0.87].

Conclusions:

A succinct logistic regression model can accurately predict severely injured pediatric patients, which could be used for prehospital trauma triage. However, there remains a critical need to obtain delta-V in real-time.

INTRODUCTION

The Centers for Disease Control (CDC) estimates that, each year, over 800,000 pediatric patients under the age of 18 years are seen in US emergency departments for injuries from motor vehicle collisions (MVCs) (CDC 2020). This makes MVCs one of the most common mechanisms of traumatic injury in pediatric patients and remains the leading cause of death in patients aged 2–18 years (CDC 2018, 2019; Cunningham et al. 2018; Tracy et al. 2013). Transporting severely injured children directly to a trauma center has been shown to decrease mortality, and pediatric-specific trauma centers may offer additional benefit (Hall et al. 1996; MacKenzie et al. 2006; Potoka et al. 2000; Walther et al. 2014; Wang et al. 2013). A recent study from France—in which MVCs were the second most common mechanism of injury—showed a six-fold increased risk of death for severely injured pediatric patients who were initially transported to a non-trauma center (Ageron et al. 2021). Given the frequency of MVCs involving children, it is essential to be able to accurately determine which pediatric patients are at risk for severe internal injury. The mechanism of injury stage of the CDC guidelines for the field triage of injured patients attempts to identify these patients (Sasser et al. 2012). However, these guidelines, as currently implemented, do not meet the American College of Surgeons (ACS) goals for determining which patients should be transported to trauma centers (Newgard et al. 2016; Rotondo 2014).

Multiple studies have demonstrated the ability of vehicle telemetry data to provide accurate predictions of severe injury in occupants involved in MVCs (Augenstein et al. 2001, 2002; Bahouth et al. 2004; Champion et al. 2004; Kononen et al. 2011; Stitzel et al. 2016). Advanced Automatic Crash Notification (AACN) is a framework in which vehicle telemetry is automatically uploaded to a central server when an impact is detected and information is distributed to key agencies within the emergency medical system (EMS) (CDC 2008; Champion et al. 1999, 2004; Champion and Cushing 1999). AACN enables the integration of vehicle telemetry data into injury predication models. A recent algorithm developed for a proprietary system outperformed existing trauma guidelines for real-world crashes (He et al. 2019). A pediatric specific injury prediction model has been developed to predict the risk of injuries requiring care at a trauma center (Doud et al. 2018; Weaver 2021).

While AACN networks have the potential to improve trauma triage, these systems have only been implemented by select vehicle manufacturers and may require a subscription service (Butler 2002; Seekins et al. 2013). There have been efforts to create triage algorithms using only information that can be obtained at the scene of a crash, known as on-scene prediction injury severity prediction (OSISP) (Buendia et al. 2015; Candefjord et al. 2015; Newgard et al. 2002; Nishimoto et al. 2019; Scheetz et al. 2007). These models show significant discriminatory power, but could benefit from the addition of information only available from telemetry data. EMS providers might be able to download and use telemetry data at the scene of a crash until AACN implementation becomes more widespread. However, simple criteria are necessary in order to use this data in the complex prehospital environment.

The objective of this study was to determine the minimum set of patient and crash variables needed to predict severe injury in pediatric MVC occupants without a significant decrease in accuracy. A secondary objective was to assess the improvement in accuracy due to the inclusion of total change in velocity (delta-V) due to the crash.

METHODS

Study Design

This study was a retrospective, population-based analysis of national crash data from 2000–2018. Based on the criteria set forth in the NIH’s common rule and the Institutional Review Board (IRB) at the University of Virginia, this study was considered exempt from IRB review since data is deidentified and publicly available.

Data Source

Data were obtained from the National Automotive Sampling System (NASS) for the years 2000–2015 and the Crash Injury Surveillance System (CISS) for the years 2017–2018. These data are collected and made available through the National Highway Traffic Safety Administration (NHTSA). Crashes which occurred more than ten years after the model year were excluded from NASS because NHTSA did not pursue data collection for these cases after 2009. Data are freely available for download from the NHTSA website (www.nhtsa.gov).

Participant Selection

Pediatric occupants, age 18 years old or younger at the time of the collision, were selected for inclusion in this analysis. Cases were excluded if there was missing data regarding: occupant age, sex, seat belt use/child restraint use, delta-V, principal direction of impact (PDOF), rollover status, posted speed limit, or injury status and death. We chose not to use imputation for these variables primary analysis because we found data was missing not at random (MNAR). However, the results of this analysis performed with multiple imputation using chained equations (MICE) can be found in Supplement 1.

Occupant characteristics

All occupant demographics were represented as categorical variables. Occupant ages were categorized as 0–4, 5–9, 10–14, and 15–18 years, based the modified version of the CDC age groups suggested by Doud et al. for pediatric injury prediction (Doud et al. 2016). Restraint use was represented with two variables: 1) any restraint use; 2) optimal restraint based on age/size-specific recommendations. Any restraint use meant there was a children restraint or manual belt in use at the time of the crash, and optimal restraint was determined based on the American Academy of Pediatrics (AAP) 2002 guidelines (Prevention 2002). Both variables were recorded as true if the occupant was optimally restrained. When patient height or weight was not available for identifying optimal restraint use, the value was imputed using the mean of known occupants of the same age and sex. A variable was collected to indicate if the occupant was in the front row versus all rear rows. Any occupant under the age of 13 years who was in the front row was considered not optimally restrained based on the AAP guidelines. There was also a variable to indicate if the occupant was ejected, which could have been a partial or complete ejection.

Crash characteristics

The total delta-V and posted speed limit were recorded as a continuous variables and all other crash variables were categorical variables. PDOFs determined were based on quadrant (frontal: 320–360 or 0–40 degrees, right: 50–130 degrees, rear: 140–220 degrees, and left: 230–310 degrees). Right and left were then recoded as near-side or far-side based on the seating position of occupant. Side impact crashes were considered far-side for middle seat occupants. Multiple collisions were derived from the number of PDOFs recorded for the crash. Rollover status was included as a binary variable. The number of quarter turns was not used, since EMS providers may not be able to accurately determine this at the scene of a crash. Any airbag deployment was true if an airbag at any position deployed as a result of the crash.

Outcomes

Two injury outcomes were used for training and evaluation. The first outcome of interest was severe injury, as determined by an Injury Severity Scale (ISS) ≥ 16 (Lee et al. 2008). Any occupant that died was considered to be positive to the outcome regardless of recorded injuries. The second outcome was if the occupant sustained any injury on a Target Injury List (TIL), developed by Weaver et al. (Stitzel et al. 2016; Weaver 2021). The TIL is contains the most common Abbreviated Injury Score (AIS) severity 2+ pediatric injuries from AIS 1990 update 98 (Association for the Advancement of Automatic Medicine 2001) based on crash mode. This list is optimized based on the time-sensitivity, predictability, and severity of injuries. Only NASS CDS was used for training and evaluation with the TIL since CISS does not contain the AIS 1990 update 98 lexicon.

These two outcomes were separately evaluated in order to measure the robustness of our method for developing a reduced-variable model. Significant differences in the importance of variables between the two outcomes would indicate instability of this technique.

Injury prediction modeling

Multivariate logistic regression was used to create injury prediction models. Since variables were not required for the reference level of the categorical predictors (age groups and PDOF), there were nineteen total predictors that were represented in seventeen variables. The receiver operator curve (ROC)-area under curve (AUC) was used to compare models. For all models, the mean AUC and 95% CI were determined using ten-fold cross validation. The folds were assigned prior to modeling and were held constant between models.

Bayesian model averaging

Bayesian model averaging is a type of Bayesian inference that can be used to determine variable importance in the setting of model uncertainty. This approach has been shown to produce more robust results than step-wise regression (George and McCulloch 1997; Madigan and Raftery 1994). Sampling methods such as Markov chain Monte Carlo are commonly used to estimate the posterior probabilities (Madigan et al. 1995). However, for our application, the relatively small number of variables allowed us to evaluate all combinations of predictors. A uniform prior was employed, so every model was considered equally likely. The derivation of the equation used to determine the posterior probabilities for each variable can be found in Supplement 2.

Analysis of minimum predictor set

The results from the complete set of multivariate logistic regression models were then used to determine the minimum number of predictors needed to maintain model accuracy. Models were grouped by the number of predictors, and within each group the model with the highest mean cross-validation estimated AUC was selected for further analysis. The distribution of AUCs obtained during cross-validation for these best performing models was compared to those of the base model which included all predictors. While the DeLong Test is often used to compare AUCs, this has been shown to be inaccurate for nested models (DeLong et al. 1988; Demler et al. 2012). Instead, we determined the mean and variance from the individual samples, then used a two-tail t-test, with a significance threshold set at p<0.05 (LeDell et al. 2015; Li 2012).

Reduced-predictor model selection and evaluation

The top k-predictors obtained from the Bayesian model averaging were then selected for the reduced model, with k representing the minimum number of predictors included in the model before there was a statistically significant drop in AUC. The reduced models were evaluated again using ten-fold cross-validation with a new set of random fold assignments.

We then estimated the performance of both models in the general population of pediatric patients in the US using the full dataset. Prediction of ISS ≥ 16 was used for both models, since this criterion is typically used as a surrogate for when a patient should be transferred to a trauma center (Lee et al. 2008). We calculated the expected under-triage rate (1-sensitivity) and over-triage rate (1-specificity) using the case weights provided in NASS and CISS. This performance was compared to the ACS goals of ≤5% under-triage (sensitivity ≥95%) and ≤35% over-triage (specificity ≥65%) (Rotondo 2014).

Models excluding Delta-V

We then evaluated the results from models that did not contain delta-V in order to examine the implications of not including this variable in the model. We followed the same process described above for determining the best performance for each number of variables. The minimum number of variables was then determined for this subset of models.

RESULTS

There was a total of 32,194 records for occupants 18 years or younger in the combined dataset, 29,788 from NASS and 2,406 from CISS. 18,541 records were excluded because of missing data (Figure 1). This resulted in 13,653 occupants for analysis, of which 644 (4.7%) were severely injured based on ISS≥16. The median age of the occupants in the study was 15 years [IQR: 7–17], with greater numbers in the oldest age groups (Table 1). The majority of occupants were restrained (77.8%), although only 53.9% were optimally restrained based on the APA guidelines. Frontal crashes were the most common PDOF (66.1%) and rollover collisions were infrequent (5.1%).

Figure 1.

Flowchart of patients included in this analysis. [NASS: National Automotive Sampling System, CISS: Crash Injury Surveillance System, Delta-V:change in velocity, PDOF: principal direction of force]

Table 1 –

Demographics for pediatric occupants analyzed in this study. All patients include occupants from NASS and CISS, used with outcome of severely injured (ISS≥16). The Target Injury List was only applied to NASS because of AIS version compatibility. [NASS: National Automotive Sampling System, CISS: Crash Injury Surveillance System, Delta-V: change in velocity, PDOF: principal direction of force, ISS: Injury severity score, AIS: Abbreviated Injury Score]

	All patients	ISSS≥16	NASS only	Target Injury List
	(n=13,653)	(n=644)	(n=12,485)	(n=680)
Age, years
Median (IQR)	15.0 (7.0–17.0)	16.0 (12.0–17.0)	15.0 (7.0–17.0)	16.0 (11.0–17.0)
0–4	2376 (17.4%)	85 (13.2%)	2149 (17.2%)	88 (12.9%)
5–9	2152 (15.8%)	46 (7.1%)	1954 (15.7%)	65 (9.6%)
10–14	2178 (16.0%)	88 (13.7%)	1971 (15.8%)	114 (16.8%)
15–18	6947 (50.9%)	425 (66.0%)	6411 (51.3%)	413 (60.7%)
Female	6783 (49.7%)	283 (43.9%)	6162 (49.4%)	304 (44.7%)
Front row	6992 (51.2%)	410 (63.7%)	6425 (51.5%)	398 (58.5%)
Any restraint	10618 (77.8%)	334 (51.9%)	9643 (77.2%)	341 (50.1%)
Optimal restraint	7356 (53.9%)	241 (37.4%)	6606 (52.9%)	246 (36.2%)
Delta-V	22.0 (15.0–31.0)	43.0 (32.0–56.0)	22.0 (16.0–31.0)	40.0 (29.0–53.0)
PDOF
Frontal	9018 (66.1%)	353 (54.8%)	8312 (66.6%)	324 (47.6%)
Rear	1220 (8.9%)	21 (3.3%)	1070 (8.6%)	42 (6.2%)
Nearside	1600 (11.7%)	172 (26.7%)	1432 (11.5%)	155 (22.8%)
Farside	1808 (13.2%)	98 (15.2%)	1671 (13.4%)	159 (23.4%)
Rollover	696 (5.1%)	91 (14.1%)	593 (4.7%)	105 (15.4%)
Multiple collisions	5121 (37.5%)	374 (58.1%)	4671 (37.4%)	413 (60.7%)
Ejection	220 (1.6%)	96 (14.9%)	196 (1.6%)	90 (13.2%

The Bayesian model averaging found that delta-V had the highest posterior probability when the outcome of interest was ISS ≥16 or the target injury list (Figure 2). The difference between probability of delta-V and next best predictor was greater than any difference between any other two predictors. The posterior probability for predictors varied slightly between outcomes. However, the order of importance of the predictors was exactly the same.

Figure 2 –

Variable importance based on Bayesian model averaging. Higher posterior probability indicates greater variable importance. [Delta-V: change in velocity, PDOF: principal direction of force]

The baseline model with all predictors was found to have a nearly mean AUC for both ISS ≥16 and TIL (ISS≥16: AUC 0.91 [95% CI: 0.86–0.95], TIL: AUC 0.90 [95% CI: 0.86–0.94]). There was no statistically significant decrease in mean AUC until the models were reduced to four or five predictors depending on the outcome (Figure 3). It was therefore determined that the minimum number of predictors necessary to predict ISS ≥16 was five, and the target injury list was six (ISS: k=5, TIL: k=6). Based on this analysis the reduced models were:

$$\begin{gathered} \mathit{l}\mathit{o}\mathit{g}\mathit{o}\mathit{d}\mathit{d}\mathit{s}\left(\mathit{P}\mathit{r}\left[\mathit{I}\mathit{S}\mathit{S}>=\mathit{16}\right]\right)={\mathit{B}}_{\mathit{I}\mathit{0}}+{\mathit{B}}_{\mathit{I}\mathit{1}}\mathit{D}\mathit{e}\mathit{l}\mathit{t}\mathit{a}-\mathit{V}+{\mathit{B}}_{\mathit{I}\mathit{2}}\mathit{E}\mathit{n}\mathit{t}\mathit{r}\mathit{a}\mathit{p}\mathit{m}\mathit{e}\mathit{n}\mathit{t}+{\mathit{B}}_{\mathit{I}\mathit{3}}\mathit{E}\mathit{j}\mathit{e}\mathit{c}\mathit{t}\mathit{i}\mathit{o}\mathit{n}+{\mathit{B}}_{\mathit{I}\mathit{4}}\mathit{A}\mathit{n}\mathit{y}\_\mathit{r}\mathit{e}\mathit{s}\mathit{t}\mathit{r}\mathit{a}\mathit{i}\mathit{n}\mathit{t}*+{\mathit{B}}_{\mathit{I}\mathit{5}}\mathit{N}\mathit{e}\mathit{a}\mathit{r}\mathit{s}\mathit{i}\mathit{d}\mathit{e} \\ \mathit{l}\mathit{o}\mathit{g}\mathit{o}\mathit{d}\mathit{d}\mathit{s}\left(\mathit{P}\mathit{r}\left[\mathit{I}\mathit{S}\mathit{S}>=\mathit{16}\right]\right)={\mathit{B}}_{\mathit{I}\mathit{0}}+{\mathit{B}}_{\mathit{I}\mathit{1}}\mathit{D}\mathit{e}\mathit{l}\mathit{t}\mathit{a}-\mathit{V}+{\mathit{B}}_{\mathit{I}\mathit{2}}\mathit{E}\mathit{n}\mathit{t}\mathit{r}\mathit{a}\mathit{p}\mathit{m}\mathit{e}\mathit{n}\mathit{t}+{\mathit{B}}_{\mathit{I}\mathit{3}}\mathit{E}\mathit{j}\mathit{e}\mathit{c}\mathit{t}\mathit{i}\mathit{o}\mathit{n}+{\mathit{B}}_{\mathit{I}\mathit{4}}\mathit{A}\mathit{n}\mathit{y}\_\mathit{r}\mathit{e}\mathit{s}\mathit{t}\mathit{r}\mathit{a}\mathit{i}\mathit{n}\mathit{t}*+{\mathit{B}}_{\mathit{I}\mathit{5}}\mathit{N}\mathit{e}\mathit{a}\mathit{r}\mathit{s}\mathit{i}\mathit{d}\mathit{e}+ \\ {\mathit{B}}_{\mathit{T}\mathit{6}}*\mathit{M}\mathit{u}\mathit{l}\mathit{t}\mathit{i}\mathit{c}\mathit{o}\mathit{l}\mathit{l} \end{gathered}$$

Figure 3 –

Maximum perform bases on the number of variables in the model. The gray bands represent the 95% confidence interval. The vertical gray lines indicate where the model performance began to be statistically significantly different that the baseline model with all predictors (p<0.05). [ISS: Injury Severity Score]

The performance based on cross-validation of the reduced-variable model for predicting ISS ≥16 was AUC 0.90 [95% CI: 0.87–0.94], and of the model predicting the target injury list was AUC 0.89 [95% CI: 0.86–0.92]. In the models that did not include delta-V, the highest AUC for predicting ISS ≥16 achieved was 0.82 [95% CI: 0.74–0.90] from a twelve-variable model. There was no statistically significant drop in performance until the model without delta-V was reduced to six variables (posted speed limit, nearside impact, restraint use, ejection, airbag deployment, entrapment).

The estimates for overtriage and undertriage in the entire pediatric population of the US involved in MVCs is reported (Table 2). The models produce a predicted probability of injury, and the threshold for the probability that triggers transport to a trauma center can be adjusted. We report the estimated rates undertriage and overtriage based on different probability thresholds.

Table 2 –

Weight estimates for undertriage and overtriage based on different probability thresholds for both outcomes. Estimates are based on the case weights provided by NHTSA. [ISS: Injury Severity Score]

Probability Threshold	ISS ≥16		Target Injury List
	Undertriage rate	Overtriage rate	Undertriage rate	Overtriage rate
0.000	0.00	1.00	0.00	1.00
0.005	0.00	0.84	0.00	0.90
0.010	0.11	0.47	0.01	0.58
0.015	0.13	0.30	0.03	0.40
0.020	0.19	0.19	0.06	0.29
0.025	0.21	0.14	0.07	0.22
0.030	0.24	0.11	0.09	0.17
0.035	0.25	0.09	0.10	0.13
0.040	0.27	0.07	0.14	0.12
0.045	0.33	0.06	0.16	0.10
0.050	0.34	0.05	0.19	0.08

DISCUSSION

Our analysis of national data of pediatric occupants involved in MVCs showed that a reduced predictor set model was effective for predicting severe injury. A parsimonious model using only five or six predictors had performance that was nearly identical to a more complicated model that used all seventeen potential predictors. These results indicate that it may be possible to create concise criteria that could be integrated into our current prehospital practices. However, obtaining an accurate injury risk hinges on the ability to obtain an estimate of delta-V at the time of EMS evaluation.

The analysis extends the results reported in many previous studies of the ability of AACN to predict injury severity (Champion et al. 1999; Doud et al. 2016; Kononen et al. 2011; Nishimoto et al. 2017; Stitzel et al. 2016; Weaver et al. 2017). AACN has been shown to be a potentially powerful tool to assist with trauma triage and has good accuracy. The AUCs reported in this study exceeds those reported in previous studies, which is likely due to our inclusion of ejection and entrapment as predictors. Ejection and entrapment are important predictors but are not typically used in AACN algorithms since they cannot be determined remotely. Since the tool we are developing is designed to be used by EMS providers at the scene of a crash, we felt the inclusion of these predictors was appropriate.

It was interesting that the predictors representing the pediatric age groups all had relatively low importance and were not included in the final models. This seems to contradict the findings reported by Doud et al., which showed the risk of injury to different body regions varied significantly between these age groups (Doud et al. 2016). However, the outcomes we examined are based on the overall injury status of the individual. The different risk levels due to age for individual body regions appear to offset, producing a relatively constant whole-body risk of severe injury.

Our results demonstrate that obtaining an accurate injury risk is significantly augmented by an estimate of delta-V at the time of EMS evaluation. The accuracy of models that did not include delta-V (AUC 0.82) were remarkably similar to those obtained by Buendia et al. for their OSISP algorithm (AUC 0.83) (Buendia et al. 2015). However, our results showed the predictive performance of these models can be improved with the addition of delta-V. Currently, almost all prehospital care triage decisions in both pediatrics and adults are made without the benefit of knowing this information. The optimal situation would be an AACN system that is installed in all vehicles and is well-integrated with emergency response systems. However, even if this technology were implemented immediately, it would likely take over a decade for this technology to penetrate the majority of fleet of vehicles on the road. Until we have that level of coverage with AACN, we should seek methods to obtain and use telemetry data given its importance. This could mean using EDR download devices that are already commercially available or developing a device specifically designed for prehospital providers.

The ability to integrate delta-V into the CDC guidelines for field triage of the injured patient has the potential to greatly improve the accuracy of trauma triage for occupants involved MVCs. Even when followed exactly, the current guidelines do not meet the ACS goals for undertriage and overtriage in pediatric patients. The CDC guidelines already include criteria for ejection and intrusion at the occupant’s site. This means that when these models are integrated into the current guidelines it is possible that ejection and near side impact could be removed. That would leave a three or four variable model which could be represented in a decision tree, chart, or scoring system. While this type of simple representation would be conducive to use in the potentially chaotic prehospital environment, the answer to the open question of whether EMS providers would use such a tool if it were available remains pivotal. A study by Newgard et al. showed that EMS provider judgement was the most commonly used triage criterion (40% all triage-positive patients). However, 60% of patients were triage-positive because of other criteria (Newgard et al. 2012). This data indicates that EMS providers are willing to use objective trauma triage criteria when they are available. Another issue is the ease of obtaining the telemetry data in the prehospital environment while AACN systems are still under development. At this time, there is only one commercially available telemetry download device that is compatible with multiple vehicle manufacturers, and further research is needed to assess the ability of EMS to use this device at the scene of a crash.

Our results indicate that we could likely improve current trauma triage accuracy by instituting a risk prediction model, even if delta-V is unavailable. When delta-V is not in the model, the posted speed limit becomes a significant predictor. The potential disadvantage is that such a model could significantly underestimate the risk for cases in which the vehicle was substantially exceeding the posted speed limit. However, a model without delta-V has the advantage that it could be implemented immediately and have moderate discriminator ability. This finding aligns with previous work that examines the use of on-scene information for trauma triage. For example, Newgard et al. demonstrated a clinical decision rule with three elements (GCS<15, intrusion ≥6 inches, or unrestrained) had a 86% sensitivity and 74% specificity for identifying pediatric patients with severe injury or need for specialized trauma care (Newgard et al. 2005).

Bayesian model averaging was used to minimize the variability of the results obtained from this analysis. Bayesian model averaging has the ability to decrease the variance in variable selection when compared to traditional methods such as stepwise and all subset regression (Genell et al. 2010; Raftery et al. 1997; Wang et al. 2004). This technique has been shown to be effective in applications such as predicting population-based behavior, modeling ecologic systems, and medical illness prediction (Andrews and Baguley 2013; Wilson et al. 2010; Wintle et al. 2003). Bayesian model averaging has found use in traffic safety, but this has previously been confined to system-level collision prediction (Blattenberger et al. 2013; Zou et al. 2012). Our study shows that these techniques can successfully be used for injury prediction.

The reduced variable models we obtained were very similar for our two injury outcomes, ISS≥16 and the target injury list. The Target Injury List is a subset of all AIS 2+ injuries selected on the basis of time-sensitivity, severity, and predictability. Models developed using the TIL have outperformed traditional AACN algorithms for predicting the need for transport to a trauma center (Stitzel et al. 2016), and we observed this performance in the results of this study. We chose to use two separate but related outcomes in order to determine the robustness of our technique. If this technique overfit the data, we would expect a significant difference in variable importance between outcomes. We observed that both outcomes produced identical order of variable importance and resulted in very similar reduced-variable models. This supports the validity of this approach for creating a parsimonious model.

These models are intended to augment the CDC field triage guidelines and give clarity to the existing vehicle telemetry criterion. We found favorable performance of our models when we examined the expected rates of under- and over-triage in the US population of pediatric occupants; however, in isolation, these models were not clearly able to meet the ACS goals. Both models were close to the ACS targets given different probability thresholds, but it is notable that the TIL model showed superior performance at every level. Further research is needed to determine what would be an optimal cut-off when the model is used in the context of the other CDC criteria.

Limitations

There are certain limitations to this study due the retrospective nature of the analysis. First, we had to exclude a significant number of cases due to missing critical data points. This has the potential to bias our results because data were not missing at random, therefore real-world validation will be critical. Second, there are limitations in the ability of ISS ≥16 to predict the need for trauma center level care, especially in pediatrics. A better metric is actual utilization of in-hospital resources which is not available in NASS or CISS. Despite these limitations, ISS is widely used as a surrogate for trauma triage evaluation. Third, the primary AIS version in which injuries were recorded changes between years. These changes in AIS versions led us to exclude certain cases, as described in the Methods. Finally, the delta-V determinations used in this analysis were obtained through crash reconstructions rather than directly from telemetry data in most circumstances. Nevertheless, these estimates have been shown to be comparable to the delta-V obtained from EDR recordings (Gabler et al. 2004; Johnson and Gabler 2014).

In conclusion, these results show that a parsimonious model can be highly accurate for predicting which pediatric patients should be directly to a trauma center after an MVC. Our findings demonstrate the critical importance of having an accurate estimation of delta-V. Future work will focus on integrating an injury prediction model into the CDC guidelines and developing a concise representation that can be used in the prehospital environment.

DATA AVAILABILITY

The authors of this manuscript are committed to transparency and reproducibility in research. The code files used for this analysis are publicly available at https://github.com/thartka/peds_mvc_triage. This code is made available under GNU General Public License v3.0. To view a copy of this license, visit https://www.gnu.org/licenses/gpl-3.0.en.html.

The data can be obtained from the National Highway Traffic Safety Administration (NHTSA) website at https/www.nhtsa.gov. Contact the Center for Injury Biomechanics at Wake Forest for information regarding the Target Injury List.