Influence of Injury Risk Thresholds on the Performance of an
Algorithm to Predict Crashes with Serious Injuries
George Bahouth
Impact Research, Inc.
KennerlyDigges
Carl Schulman
The William Lehman Injury Research Center
__________________________________
ABSTRACT –
This paper presents methods to estimate crash injury risk based on crash characteristics captured by some
passenger vehicles equipped with Advanced Automatic Crash Notification technology. The resulting injury risk estimates could
be used within an algorithm to optimize rescue care. Regression analysis was applied to the National Automotive Sampling
System / Crashworthiness Data System (NASS/CDS) to determine how variations in a specific injury risk threshold would
influence the accuracy of predicting crashes with serious injuries. The recommended thresholds for classifying crashes with
severe injuries are 0.10 for frontal crashes and 0.05 for side crashes. The regression analysis of NASS/CDS indicates that these
thresholds will provide sensitivity above 0.67 while maintaining a positive predictive value in the range of 0.20.
__________________________________
INTRODUCTION
Automatic Crash Notification Systems are now
capable of transmitting vehicle data, in addition to
geographic coordinates, after a crash with air bag
deployment. In 2008, an expert panel sponsored by
CDC recommended the transmission of the
following crash variables: delta V, crash direction,
belt use, multiple impact and rollover crashes [CDC
2008]. This data would be used as input to an
algorithm that could predict those crashes most
likely to require urgent response. The panel
recommended that the algorithm prediction should
not cause an over-triage (false positive) greater than
4 out of 5. This false positive ratio was expressed in
the CDC report as a 20% risk of serious injury. The
20% requirement would limit the positive predictive
value (PPV) of the algorithm to a value greater than
0.20.
The use of algorithms that use crash data to predict
injury risk was the subject of research sponsored by
the National Highway Safety Administration in
1996. One of the resulting papers showed the
influence of as many as 23 crash variables on injury
risk [Malliaris, 1997]. In the final presentation to
NHTSA on March 27, 1997, the most influential
variables from this analysis were programmed to
provide a graphic depiction of the injury risk for any
combination of values for the crash variables. This
algorithm was named URGENCY and was then
applied by NHTSA in their field trial of Automatic
Crash Notification Systems [Kanianthra, 2000].
The application and benefits of the system to predict
crash injuries were articulated in a paper written by
the participants in the NHTSA study [Champion,
1999]. A more focused study separated the crash
modes and evaluated the improved accuracy for
each added variable [Bahouth, 2003].
Earlier papers [Malliaris, 1997, Augenstein, 2006]
summarized the methodology for predicting injury
risk from crash data that could be transmitted by an
automatic crash notification system. The crash data
for the algorithm was based on variables in the
NASS/CDS database for the years 1997 to 2003.
More recently, the Center for Disease Control
(CDC) published a report [CDC 2008] that
summarized recommendations from an expert panel
for the crash data to be transmitted by ACN
systems. The CDC report did not address the
interpretation of the transmitted data. A goal of this
study is to further develop and evaluate predictive
algorithms from the data elements recommended in
the CDC report, applying the methodology
employed in earlier research [Malliaris, 1997,
Bahouth, 2003, Augenstein, 2006]. Additional
years of NASS/CDS data are now available to
enrich the database over that used in the earlier
papers.
5
6
th
AAAM Annual Conference
AnnalsofAdvancesinAutomotiveMedicine
October
14‐17,201
2
Vol56 • October2012
223
Attachment (18)
A significant knowledge gap exists in determining
the injury risk threshold required to capture a large
percentage of vehicle occupants with serious
injuries without overloading the response system
with those who do not need urgent medical
attention. There is no previously published research
to address this problem. A principal objective of
this paper is to determine the risk thresholdsof a
NASS based algorithm that will correctly identify a
large fraction of the seriously injured while
providing a positive predictive value (PPV) greater
than 0.20.
METHODS
The National Automotive Sampling
System/Crashworthiness (NASS/CDS) used in this
study is the only available database that provides
detailed information on injuries and crash factors
and is representative of the crashes in a geographic
area – in this case the US. The requirement for
entry into NASS/CDS is that one of the vehicles
involved in the crash must have been damaged
sufficiently to be towed away from the scene. The
NASS/CDS is a stratified sample, with the more
severe crashes being sampled more frequently than
the less severe crashes. Each case is assigned a
weighting factor so that each case can be
extrapolated to estimate the frequency of various
crashes and injuries in the United States. Within
NASS/CDS, specific injuries sustained, including
their severities, are recorded allowing for the direct
association of crash conditions with crash outcomes
as used in this study.
This study addresses passenger vehicle occupants
over the age of 16 who may have severe or time
critical injuries following a crash. This category
includes occupants who sustained at least one or
more injuries with an Abbreviated Injury Severity
(AIS) Score of 3 or those who were fatally injured
during a crash due to trauma. AIS 3 or higher
injuries are serious (AIS3), severe (AIS4), critical
(AIS5) and maximum (AIS6) injuries. Throughout
this text, these occupants will be referred to as
MAIS 3+F injured occupants.In NASS/CDS
weighted data, approximately 3.5 million front seat
occupants are exposed to crashes each year (1998-
2007). Of this population, 101,000 are injured with
a Maximum AIS score of 3 or higher injuries (MAIS
3+F).
The injury predictive algorithm was developed
using multiple regression analysis of the crash
variables recommended by the CDC panel.
Separate algorithms were developed for different
crash directions. The algorithms were trained using
NASS 1998-2007. NASS 2008 and 2009 were used
to test its predictive accuracy of the algorithms. The
predictive accuracies were found to vary with the
risk threshold applied to the algorithm. The risk
thresholds were varied and the resulting predictions
were evaluated to determine the threshold that
corresponded to a PPV of 0.20.
Findings from available literature confirm the use of
change of velocity (deltaV) as a measurement of
crash energy and crash severity. (Malliaris 1997,
Jones 1989, Siegel 1993, Augenstein 2003). There
is, however, an issue with the crash severity
predictions in NASS. The deltaV is based on a
calculation of the energy absorbed by the vehicle
structure. It relies on measurements of the vehicle
damage and estimates of the masses of the vehicles,
their stiffness and the direction of the crash.
Slightly more than half of the cases of the
approximate severity to deploy the airbag have a
deltaV recorded. For the purpose of this study, it is
assumed that the crashes without deltaV recorded
have the same crash severity as those with recorded
deltaV and their exclusion will not likely bias
estimates presented here.
The ability to manage the kinetic energy of a
vehicle and occupant depends largely on the primary
direction that decelerating forces are applied. For
example, in frontal crashes the frontal crush zones,
seatbelts and frontal airbag systems help to manage
energy along the longitudinal axis of the vehicle.
These features are less effective in reducing injury
to nearside occupants in lateral crashes of equivalent
severity as measured by deltaV.
For this study, crash mode has been categorized
using Collision Deformation Classification (CDC)
data collected by NASS/CDS investigators. Each
mode is categorized as follows:
Frontal: Any Seating Position,(PDOF≥11 and
PDOF≤1), or (PDOF=10 or 2 where General Area
of Damage is Front)
Nearside: (PDOF≥2 and PDOF≤4, Right Seating
Position, General Area of Damage is Right) or
(PDOF≥8 and PDOF
≤10 and, Left Seating Position,
General Area of Damage is Left)
Farside: (PDOF≥2 and PDOF≤4, Left or Middle
Seating Position, General Area of Damage is Right)
or (PDOF≥8 and PDOF≤10 and, Right or Middle
Seating Position, General Area of Damage is Left)
Rear: PDOF≥ 5and PDOF≤7
©
AnnalsofAdvancesinAutomotiveMedicine
224
These same crash categories for frontal and side
impacts were published and applied by NHTSA
during the Final Economic assessment of the
FMVSS Advanced Airbag Final Rule (NHTSA,
2000). In the regression analysis to follow, these
definitions will be applied to occupants seated at the
driver and right front passenger positions.
Regression Analysis
The application of crash variables to the population
of occupants in crashes above the ACN threshold
attempts to identify front seat occupants who may
be seriously to fatally injured. A model which
estimates injury risk based on crash characteristics
can be applied. This model or approach to
processing crash information to improve rescue care
is known as the URGENCY algorithm [Malliaris,
1997]. The higher the injury risk, the larger the
expected proportion of seriously injured. If an
injury risk threshold is established and all crashes
above that risk value are designated as serious, the
resulting population will contain both seriously
injured and non-seriously injured people. Within
that population, those that are not seriously injured
are called false positives. The proportion of
seriously injured occupants who are correctly
identified as injured by the model is known as
sensitivity. The population that is designated as not
serious crashes will also contain some seriously
injured people. These seriously injured people that
are missed are called false negatives. The proportion
of non-seriously injured occupants who are correctly
identified as non-seriously injured is known as
specificity. The ratio of the correctly identified
MAIS 3+ to the false positives plus the true
positives is known as the positive predictive value
(PPV). The challenge is to select a risk threshold
that does not contain excessive false positives, yet
does not allow an excessive number of false
negatives. The CDC has proposed that a PPV of .20
is acceptable for triage decisions for transporting
injured people to a trauma center [CDC 2008].
An objective of the regression analysis was to
determine appropriate injury risk thresholds for the
different crash directions. A principal constraint on
the analysis was the requirement that the PPV value
not be less than .20.
The variables identified by the CDC Committee for
transmission with the ACN data were as follows:
- Delta V
- Principal direction of force (PDOF)
- Seatbelt usage/or without
- Crash with multiple impacts
- Vehicle type
In addition, the Committee recommended that the
voice communication with crash involved occupants
determine if any were 55 years old or older. For the
purposes of this analysis, rollover after a front or
side crash was added as a class of multiple impacts.
In order to take into account multiple factors
influencing crash severity and the likelihood of
injury, multiple regression techniques were used.
Since the outcome of interest could fall into one of
two categories (MAIS 3+ injured or non-MAIS 3+
injured), binary logistic regression is ideally suited
for the analysis. In addition, certain high severity
crash attributes like the occurrence of complete
occupant ejection were assumed to indicate high
probability of severe injury even in the absence of
other crash factors.
Binary logistic regression relates the contribution of
independent predictor variables (crash conditions)
with dependant outcomes (injury). Using the
Principle of Maximum Likelihood, an estimate of
the likelihood of the outcome (injury) is derived on
a scale from 0 to 100% probability. The method is
described in detail in earlier publications [Malliaris
1997, Augenstein 2003, Bahouth 2003].
Equations 1-2 show the mathematical relationship
between crash characteristics and injury outcome
probability following logistic regression model
creation. The regression parameters including the
Intercept, β1, β2… shown below are based on a least
squares fit of existing historical crash data from
NASS/CDS.
Eq. 1:
221
**)( factordeltaVInterceptw
β
β
+
+
=
Eq. 2:
))exp(1(
1
)3(
w
MAISP
−+
=+
Each logistic regression model was trained using
NASS/CDS 1998-2007 data. 2008 and 2009
datasets were used to evaluate the accuracy of the
resulting models. Initial sensitivity evaluations were
conducted using a risk prediction threshold of 0.1
for frontal and rollover crashes and 0.05 for the
other crash modes. Ultimately, assessments of
varying thresholds were conducted to determine best
risk value to produce a PPV greater than 0.2.
Before the creation of each logistic regression
model, all relevant crash attributes were reviewed
for consistency and reconditioned when appropriate
using SAS version 9.2. Analysis of the
Vol56 • October2012
225
NASS/CDSNASS CDS data was performed using
procedures appropriate for the analysis of survey
data and the correct interpretation of sample
variances for multi-stage, clustered samples.
As previously mentioned, the binary outcome
variable MAIS 3+F was used in the analysis to
distinguish injured from non-injured. For this study
MAIS 7 were considered unknown unless a fatality
occurred. These occupants were discarded from the
analysis. Cases with missing values for any model
variable are unusable for model training and testing
and were therefore discarded, as well.
Each vehicle front seat occupant was analyzed
separately. Multiple front seat occupant vehicles
would be represented twice and, in the case of
lateral collisions, would be classified as nearside
and farside crash involved.
A separate regression analysis was performed for
each crash mode. In each of these analyses, the
dependent variable was binary with MAIS 3+F
injured as 1 and all others as 0.
For frontal, near-side and far-side crash modes, the
variables used for regression were as follows:
-DeltaV – continuous in MPH
- Impact – binary – multi-impact yes or no
- Rollover - binary – rollover occurrence yes or no
- Belt Use – binary – belts used yes or no
- Age – 16-45 reference; 55-74 and 75+ groupings
For the binary variables 1 is yes and 0 is no.
The injury reduction benefits of air bags are well
established. In virtually all front and side crashes
that result in an ACN signal, the air bags will have
deployed. It is not necessary to include air bag
deployment as an injury predictor.
RESULTS
Table 1 summarizes the data from NASS/CDS
1998-2009. The data is for front seat occupants
older than 16 years. The 11 years of data are
averaged to provide annual estimates. The vehicles
are limited to model year 1998 and later. The
estimates are based on weighted data. In Table 1, the
NASS/CDS data is separated by crash mode. The
rollover category is based on the NASS rollover
category. However, rollovers with damage areas
associated with front, side or rear impacts prior to
the rollover were excluded. For these pure
rollovers, there is no basis for coding the deltaV.
Consequently, these crashes were excluded from the
algorithm analysis, but were included in Table 1.
The annual numbers of occupants with and without
MAIS 3+F injured occupants are displayed by crash
direction in Table 1.
Table 1 shows that the population of seriously
injured in a rear crash is small. It comprises only
2% of the seriously injured population. The rear
crash mode data is insufficient to generate suitable
algorithms. Consequently, this crash mode was
excluded from the regression analysis results
reported in the tables to follow.
Table1. Annual Numbers of Front Seat
Occupants (>16y/o) with and without MAIS 3+F
Injured in NASS 1998-2009
Crash
Mode
NotMAIS
3+F
MAIS
3+F
All
Frontal 4,659,399 100,404 4,759,803
NearSide 1,123,509 45,027 1,168,536
FarSide 1,127,373 21,543 1,148,916
Rear 797,255 3,572 800,827
Rollover 322,693 24,777 347,470
The pure rollovers constitute about 13% of the
seriously injured occupants. For these cases there
are no recorded deltaV and insufficient vehicle
information in NASS to generate risk factors that
are currently measured in vehicles. Development of
risk factors for pure rollovers may need to wait for
the information from rollover sensors to be collected
in real world crashes. However, another 13% of the
MAIS 3+F injured are in rollovers with a pre-roll
crash and an associated deltaV. These rollovers are
included as multiple impacts within crashes with
front or side damage.
For the exposed population in NASS 1998-2009, the
average belt use rate was 90%. For the seriously
injured it was 61%. The vehicles in NASS 1998-
2009 population were 62% cars, 19% SUV’s, 7%
vans and 13% pickups.
The baselinemodel used all the variables listed
above except age in the initial regression analysis.
The model coefficients are shown in Table 2.
©
AnnalsofAdvancesinAutomotiveMedicine
226
Table 2. Coefficients for URGENCY Algorithm
Variable
(p‐value)
Frontal NearSide FarSide
Intercept ‐5.2321
(p=<.001)
‐5.9529
(p=<.001)
‐5.0963
(p=<.001)
Delta_v 0.1335
(p=<.001)
0.2092
(p=<.001)
0.1641
(p=<.001)
Impact 0.2743
(P=0.02)
1.0401
(p=<.001)
0.6528
(p=.11)
Rollover 1.526
(p=.001)
‐0.4525
(p=<.001)
0.8247
(p=.74)
Belt_Use ‐1.1045
(p=<.001)
‐0.5558
(P=<.01)
‐1.9522
(P=<.001)
The consequences of varying probability cut points
for the frontal, near side and far side algorithms are
shown in Tables 3, 4 and 5. The prevalence of the
injuries for each crash mode are listed in Table 1.
Table 3. Sensitivity,Specificity and PPVof
Frontal Algorithm Prediction to Probability Cut
Point
Frontal
probability
cutpoint
Sensitivity Specificity PPV
0.05 0.595 0.938 0.183
0.10 0.509 0.971 0.288
0.15 0.321 0.982 0.293
0.20 0.294 0.986 0.325
Table 4. Sensitivity,Specificity and PPV of Near
Side Algorithm Prediction to Probability Cut
Point
NearSide
probability
cutpoint
Sensitivity Specificity PPV
0.05 0.674 0.900 0.175
0.10 0.511 0.954 0.257
0.15 0.436 0.975 0.351
0.20 0.373 0.984 0.427
Table 5. Sensitivity,Specificity and PPV of Far
Side Algorithm Prediction to Probability Cut
Point
FarSide
probability
cutpoint
Sensitivity Specificity PPV
0.05 0.725 0.951 0.235
0.10 0.421 0.975 0.259
0.15 0.349 0.984 0.311
0.20 0.337 0.987 0.341
In order to achieve a PPV of 0.20 or greater, the data
from the above tables suggests a cut point for frontal
crashes at 0.10 and for side crashes at 0.05. When
these cut points are applied and the age of the
occupant is included, the resulting sensitivity is
displayed in Table 6.
In a multi-vehicle collision the algorithm may be
useful in predicting injuries in other vehicles. To
investigate this possibility, the regression analysis
was applied at the accident level rather than the
occupant level. The results are displayed in Table 7.
Table 6. Sensitivity,Specificity and PPV of
Algorithm Prediction when Occupant Age is
Included
AgeInclusion
prob.cutpoint
Sensitivity Specificity PPV
Frontal(0.1) 0.594 0.882 0.269
Nearside(0.05) 0.839 0.741 0.192
Farside(0.05) 0.738 0.905 0.208
Table 7. Sensitivity,Specificity and PPV of
Algorithm Prediction when Accident Levelis
Included
A
ccident Level
p
rob. cut point)
Sensitivity Specificity PPV
Frontal (0.1)
0.679 0.524 0.202
Nearside (0.05)
0.716 0.739 0.237
Farside (0.05)
0.520 0.895 0.283
Vol56 • October2012
227
DISCUSSION
According to a recent CDC expert panel, rapid
identification and treatment of injured occupants
will improve injury outcomes and reduce deaths
following a crash [CDC 2008]. A goal of this study
was to quantify the frequency that an URGENCY
algorithm would accurately distinguish occupants
who need immediate medical attention from those
who do not. A number of previous studies have
applied the methodology developed by Malliaris and
expanded by Bahouth [Malliaris, 1997; Bahouth,
2003]. However, there have been no published
values for the coefficients used in the predictive
algorithm since the study by Bahouth in 2003. This
paper presents the URGENCY coefficients, based
on the most recent years of NASS data.
A critical impediment to the use of vehicle crash
data to improve triage decisions is the lack of an
agreed upon threshold to be used with the predictive
algorithm to trigger a rapid emergency response. If
the threshold is too low, the rescue system may be
overwhelmed by responding to crashes with minor
or no injuries. If the threshold is too high, too many
of the seriously injured may be missed. It is
essential that a threshold be agreed upon that will
capture most of the seriously injured without
misidentifying an excessive number of crashes that
do not require immediate response. The goal
established by CDC was a PPV greater than 0.2.
The discussion to follow examines the threshold and
the specificity achieved when PPV goal is achieved.
An excessive number of false positives predicted by
the injury risk algorithm could result in wasted
resources and ultimately in ignoring its prediction
by the services that are adversely affected. Several
features can act to reduce the consequence of false
positives predicted by the algorithm. The most
useful feature is the voice communication with the
occupants of the crashed vehicle that can be used to
reinforce or revoke the severe injury prediction.
The study defines serious injuries as MAIS 3+ and
fatalities. This MAIS 3+ definition is frequently
used in safety regulatory analyses [NHTSA, 2000].
An alternative definition is the ISS scale that is more
appealing to CDC [CDC 2008]. This scale
recognizes multiple injuries by summing the square
of the three highest AIS injuries. Both scales lack
precision in identifying the entire population of time
critical injuries. The URGENCY algorithm
predictive value for the MAIS scale is a percentage
risk, while the value for the ISS is a scalar number.
Since NASS is better suited to the analysis of MAIS
3+F injuries and the prediction of risk is more
intuitively appealing, that definition of serious
injury was chosen over the ISS scale.
Occupant age was found to be an influential variable
in determining injury risk in this study and in the
previous studies already cited. At present, occupant
age is not transmitted with the automatic crash
notification data. However, voice communication
with vehicle occupants can, in many cases,
determine the age range of the occupants.
Alternatively, occupant age range may be inferred
from owner registration data. In future crash
notification systems, occupant age and medical data
may be transmitted along with the geographic
coordinates and the crash data. For these reasons,
occupant age was included as a variable in this
study.
These resultsapply to front seat occupants only.
Two important factors preclude the analysis of rear
seat occupants in the same way that front seat
occupants were analyzed. First, current vehicles do
not automatically detect rear seat occupancy and
therefore could not accurately trigger an injury risk
calculation. Second, the number of seriously injured
rear seat occupants within the NASS CDS was
insufficient to train such models in order to produce
reliable estimates. Therefore, rear seat occupants
were excluded from this evaluation.
The best prediction was for far-side crashes. A risk
threshold of 0.05 produced a specificity of 0.725
and a PPV of 0.235. For near-side crashes, the same
threshold gave a specificity of 0.674 but a PPV of
0.175. For frontal crashes the best prediction was
achieved by estimating the risk of injury at the
accident level rather than the occupant level. For
that prediction, a threshold of 0.10 produced a
specificity of 0.679 and a PPV of .203.
The lowest sensitivity was for frontal crashes.
However, it was observed that frontal crashes
frequently involved crashes with the side of other
vehicles. In these crashes the occupants involved in
the side crashes were more likely to be seriously
injured. When the algorithm for frontal crashes was
applied to a dataset at the crash level rather than the
vehicle level, improved sensitivity resulted. The
interpretation of this result was that some occupants
involved in the crash would prompt medical
attention, even though that occupant was not in the
vehicle involved in the frontal crash. The sensitivity
for the frontal model applied at the accident level
was 0.679. The PPV was 0.203.
©
AnnalsofAdvancesinAutomotiveMedicine
228
The use of a risk threshold of 0.20 was considered
by the CDC Committee as an appropriate level for
classifying crashes with severe injuries. However,
Tables 3, 4 and 5 show that for that threshold, the
sensitivity ranges from 0.294 for frontal to 0.373 for
near side. A higher sensitivity would be desirable.
For the thresholds suggested in this paper,
sensitivities above 0.67 appears to be possible while
maintainintg a PPV in a range acceptable to rescue
and triage services. It should be noted that due to
the overwhelming percentage of non-seriously
injured occupants within the crash population,
higher positive predictive values could be achieved
by simply raising injury risk thresholds to the 20%
level or beyond. However, while this would
effectively improve model performance, it would
render this potentially lifesaving technology less
effective in identifying injured occupants in need of
care. Such a tradeoff must be considered when
policy decisions regarding injury risk threshold are
put forth.
The prediction of serious injuries in rollover crashes
remains a challenge. The present crash notification
systems do not transmit crash severity measures for
the rollover phase of a crash. The current injury
prediction of rollovers is based primarily on the
planar deltaV that may occur before the rollover
takes place. The presence of a rollover in
conjunction with a planar deltaV modifies the injury
risk by a fixed percent, regardless of the severity of
the rollover. It is interesting to note from Table 2
that a rollover reduces injury risk when it follows a
near side crash. An explanation may be that less of
the crash energy is absorbed by the occupant
compartment.
Another concern is the degree to which the NASS
database represents the current automatic crash
notification fleet. A recent study of NASS has
shown that injury risks are reducing in all crash
modes for the most recent vehicle model years
[Eigen, 2012]. This result suggests that an injury
predictive algorithm using the NASS database of the
current fleet would over predict serious injuries.
This over prediction would be offset to some extent
by the injuries that occur in the vehicle without
automatic crash notification when impacted by a
newer vehicle. In order to reduce the influence of
vehicles with older safety technology, model years
prior to 1998 were excluded from the analysis.
CONCLUSIONS
The recommended thresholds for classifying crashes
with severe injuries are 0.10 for frontal crashes and
0.05 for side crashes. The regression analysis of
NASS/CDS indicates that these thresholds will
provide a sensitivity above 0.67 while maintaining a
PPV in the range of 0.20.
ACKNOWLEDGEMENT
The authors would like to acknowledge the US
Centers for Disease Control for sponsoring this
research.
BIBLIOGRAPHY
Augenstein JS, Digges KH, Ogata S, Perdeck E,
Stratton J, “Development and Validation of the
Urgency Algorithm to Predict Compelling Injuries,”
Proceedings of the 17th International Technical
Conference on the Enhanced Safety of Vehicles
(ESV), June 4-7, 2001, Amsterdam, DOT HS 809
220.
Augenstein, J, Digges, K., Bahouth, G. and others,
“Characteristics of Crashes that Increase the Risk of
Injury”, 47th Annual Proceedings of the Association
for the Advancement of Automotive Medicine, p.
561-576, September, 2003
Augenstein, J, Perdeck, E., Digges, K., Bahouth, G.,
Baur, P., and Borchers. N., “A More Effective Post-
Crash Safety Feature to Improve the Medical
Outcome of Injured Occupants”’ Society of
Automotive Engineers World Congress (SAE Paper
2006–01–0457): Detroit, MI, 2006.
Bahouth, GT, “Development and Validation of
Injury Predicting Algorithms for Automotive
Applications”, George Washington University
School of Engineering and Applied Science,
Doctoral Dissertation, May 2003.
CDC 2008 “Advanced Automatic Collision
Notification and Triage of the Injured Patient”,
http://www.cdc.gov/injuryresponse/pdf/AACN%20
Report_FINAL-a.pdf.
Champion HR, Augenstein JS, Cushing B, Digges
KH, Hunt RC, Larkin R, Malliaris AC, Sacco WJ,
Siegel JH, "Urgency for a Safer America," AirMed,
March/April 1999, pp.18-23.
Collision Notification System,” SAE Proceedings,
Convergence 2000, Oct. 1, 2000, Detroit, 2000-01-
C041.
Eigen, AM, Digges, KH, and Samaha, RR, “Safety
Changes in the US Vehicle Fleet since Model Year
1990, Based on NASS Data”,Annual Proceedings of
Vol56 • October2012
229
the Association for the Advancement of Automotive
Medicine, Vol. 56, 2012.
Jones IS, Champion HR, "Trauma Triage: Vehicle
Damage as an Estimate of Injury Severity," Journal
of Trauma, Vol. 29, No. 5, May 1989, p. 646.
Kanianthra J, Carter A, Preziotti G, “Field
Operational Test Results of an Automated Collision
Notification System,” SAE Proceedings,
Convergence 2000, Oct. 1, 2000, Detroit, 2000-01-
C041.
Malliaris AC, Digges KH, DeBlois JH,
"Relationships Between Crash Casualties and Crash
Attributes," Society of Automotive Engineers SAE
970393, February 1997.
NHTSA Office of Regulatory Analysis &
Evaluation Plans and Policy “Final Economic
Assessment, FMVSS No. 208: Advanced Air Bags”,
NHTSA Technical Review, Washington, DC, May,
2000.
Siegel JH, Mason-Gonzales S, Dischinger P,
Cushing B, Read K, Robinson R, Smialek J,
Heatfield B, Hill W, Bents F, Jackson J, Livingston
D, Clark C, “Safety Belt Restraints and
Compartment Intrusions in Frontal and Lateral
Motor Vehicle Crashes: Mechanisms of Injuries,
Complications, and Acute Care Costs,” Journal of
Trauma, Vol. 5, No. 34, 1993.
©
AnnalsofAdvancesinAutomotiveMedicine
230
