Content uploaded by Joseph Roland Keebler
Author content
All content in this area was uploaded by Joseph Roland Keebler on Sep 16, 2021
Content may be subject to copyright.
IP: 192.168.39.211 On: Thu, 16 Sep 2021 16:24:19
Copyright: Aerospace Medical Association
Delivered by Ingenta
AEROSPACE MEDICINE AND HUMAN PERFORMANCE Vol. 92, No. 7 July 2021 563
RESEARCH ARTICLE
A Bayesian Approach on Investigating Helicopter
Emergency Medical Fatal Accidents
Richard J. Simonson; Joseph R. Keebler; Alex Chaparro
INTRODUCTION: Helicopter Emergency Medical Service (HEMS) is a mode of transportation designed to expedite the transport of a
patient. Compared to other modes of emergency transport and other areas of aviation, historically HEMS has had the
highest accident-related fatality rates. Analysis of these accident data has revealed factors associated with an increased
likelihood of accident-based fatalities. Here we report the results of an analysis on the likelihood of a fatality based on
various factors as a result of a HEMS accident, employing a Bayesian framework.
METHODS: A retrospective study was conducted using data extracted from the NTSB aviation accident database from April 31,
2005, to April 26, 2018. Evidence from Baker et al. (2006) was also used as prior information spanning from January 1,
1983, to April 30, 2005.
RESULTS: A Bayesian logistic regression was implemented using the prior information and current data to calculate a posterior
distribution condence interval of possible values in predicting accident fatality. The results of the model indicate that
ying at night (OR 3.06; 95% C.I 2.14, 4.48; PoD 100%), ying under Instrument Flight Rules (OR 7.54; 95% C.I 3.94, 14.44;
PoD 100%), and post-crash res (OR 18.73; 95% C.I 10.07, 34.12; PoD 100%) signicantly contributed to the higher
likelihood of a fatality.
CONCLUSION: Our results provide a comprehensive analysis of the most inuential factors associated with an increased likelihood of
a fatal accident occurring. We found that over the past 35 yr these factors were consistently associated with a higher
likelihood of a fatality occurring.
KEYWORDS: HEMS, Bayesian, accidents.
Simonson RJ, Keebler JR, Chaparro A. A Bayesian approach on investigating helicopter emergency medical fatal accidents. Aerosp Med Hum
Perform. 2021; 92(7):563–569.
Helicopter Emergency Medical Services (HEMS) are a
mode of transportation used to expedite the transport
of a patient to a care facility. The inception of this pro-
gram has its roots in the Vietnam war, in which the helicopter
was a popular method of rapidly transporting the injured from
rough or otherwise inaccessible terrain to field hospitals.17 Since
then, the use of helicopters in emergency transport has seen a
steady increase in popularity from under 25,000 EMS helicopter
flight hours in 1980 to nearly 600,000 flights hours in 2017.6,7
While the use of this technology for rapid transit is shown to
decrease patient mortality rates,16 controversy about the neces-
sity of HEMS flights has been a contentious issue for decades.
Specifically, medical professional consensus states that provid-
ing care within an hour of injury to trauma patients significant-
ly increases their chances of survival. However, a review of the
literature presents “little scientific evidence” that supports this
position.13 A review of the fatality and accident rates within the
HEMS field reveals that EMS helicopters have nearly twice as
many fatal accidents per 100,000 flight hours as any other form
of aviation.6 Although accident-related factors are associated
with an increase in fatalities in HEMS accidents,1 there is very
little evidence of the factors that contribute to increased acci-
dent rates in HEMS. In other words, we know if lives will be lost
due to a certain accident, but we have a poor understanding of
the factors that are associated with an increased risk of an
From Embry-Riddle Aeronautical University, Daytona Beach, FL.
This manuscript was received for review in September 2019. It was accepted for
publication in March 2021.
Address correspondence to: Joseph R. Keebler, Ph.D., 1 Aerospace Blvd, Daytona
Beach, FL 32114; keeblerj@erau.edu.
Reprint and copyright© by the Aerospace Medical Association, Alexandria, VA.
DOI: https://doi.org/10.3357/AMHP.5523.2021
IP: 192.168.39.211 On: Thu, 16 Sep 2021 16:24:19
Copyright: Aerospace Medical Association
Delivered by Ingenta
BAYESIAN ANALYSIS OF HEMS—Simonson et al.
564 AEROSPACE MEDICINE AND HUMAN PERFORMANCE Vol. 92, No. 7 July 2021
accident occurring. This paper intends to aggregate the past 35
yr of data on HEMS accident fatalities utilizing a Bayesian
framework to best understand the conditions that are linked to
loss of life in these accidents.
A number of studies have investigated and reviewed the state
of HEMS and its risk factors;1,3,4 however, the most recent of
these comprehensive studies reviewing HEMS fatalities in the
United States was published over a decade ago.6 Our contribu-
tions in this paper include a review of HEMS accidents using
data from 2005 to 2018 that is combined with previous research
to strengthen the evidence of what contributes to an increased
fatality likelihood. We achieve this by implementing prior evi-
dence of the factors that contribute to HEMS accidents into a
model with current data via the Bayesian framework. Due to the
lack of research and evidence of any factors other than fatal
HEMS outcomes, we are limited to what we can infer from prior
information. As such, the scope of this paper focuses solely on
what may affect the likelihood that any fatality occurs as the re-
sult of an accident by updating past evidence with current data.
Bayesian methodology is a framework that uses conditional
probability and prior information on the probability of a condi-
tion occurring to inform and estimate the credibility that an
event will occur or not. Essentially, it is a method of updating
the credibility of an event based on previous information about
said event. This framework is often expressed as its equation
(Eq. 1). The left side of the equation represents the posterior
probability or distribution, which is the updated probabilities
when new information is merged with prior data. The right side
of the equation represents the combination of the likelihood, or
current data, and the prior information.
Likelihood Parameter } Posterior Parameter 3 Prior Parameter
Eq. 1
Below we discuss a few key differences of the Bayesian meth-
odology that underlie its application in this study. First, and
mainly, Bayesian methods do not utilize P-values to interpret
significance of a viable model. Rather, various methods of iden-
tifying a variable influence on a model’s ability to predict data
are used. For the model developed in this paper, we identified
any posterior that included zero in its 95% confidence interval
and excluded it. In our application of Bayesian methodology,
we use a combination of coefficient effects and their confidence
intervals along with a priori theory to inform us of variable and
model selection as well as hierarchical model comparison via a
leave-one-out method of cross-validation.
The use of Bayesian methodology to update evidence is a
popular technique applied in a multitude of fields to improve
inferential power from the analyses conducted. For example,
Solomon and King18 presented a Bayesian application to study-
ing the factors that are associated with accident rates for fire
vehicles, and Miranda15 applied a Bayesian framework to better
understand Naval aviation mishaps. Similarly to these studies,
we use prior information about the factors associated with
higher likelihood of fatalities in HEMS accidents to update cur-
rent data.
METHODS
A retrospective study of fatalities on HEMS related flights from
April 31, 2005, to April 26, 2018, was conducted using informa-
tion available from the National Transportation Safety Board
(NTSB) aviation accident database. The inclusion criterion for an
accident was that a helicopter designated for medical duties was
involved in an unintended impact to any part of the aircraft or
any unintended incident that led to the necessity of an NTSB in-
vestigation. A total of 131 HEMS accidents were extracted from
the NTSB database. All information relevant to the accident
reported in the NTSB’s final report document was collected,
organized, and concatenated for analysis. The prior distributions
for the parameters of our model were based on previous research
identifying the key factors that contribute to the likelihood of a
fatality for HEMS accidents occurring between January 1, 1983,
and April 30, 2005.1
Procedure
Bayesian inference is a method that incorporates both prior and
current information into a model aimed at understanding condi-
tional probabilities, i.e., the likelihood of two or more events
happening together. This is accomplished using three main
parameters of a model called the prior, likelihood, and posterior.
The prior, or prior distribution, is the information from prior be-
liefs that is incorporated into the new model, or in other words,
the currently known conditional probabilities between two
events. After the current data, or likelihood, is entered into the
model, we are given the posterior distribution, which is the distri-
bution when the likelihood and prior parameters are combined.
Each factor entered into a model (e.g., all input variables) is bro-
ken down into its parameters (i.e., mean, variance, etc.), which are
assigned prior distributions which leads to a posterior distribu-
tion that contributes to the posterior prediction of the outcome.
These aspects of Bayesian analysis are summarized in a graphical
representation of Bayes’ Formula (Fig. 1).
Deciding on the most appropriate prior is an important
component of Bayesian statistics and generally receives the
Fig. 1. Graphical representation of Bayes Theorem.
IP: 192.168.39.211 On: Thu, 16 Sep 2021 16:24:19
Copyright: Aerospace Medical Association
Delivered by Ingenta
BAYESIAN ANALYSIS OF HEMS—Simonson et al.
AEROSPACE MEDICINE AND HUMAN PERFORMANCE Vol. 92, No. 7 July 2021 565
most scrutiny as a poorly formed prior can have as much influ-
ence on the outcome of the model as a diligently constructed
one. As such, the use of three types of priors have been suggest-
ed: noninformed, weakly-informed, and informed. These are
described below in order of least to most desirable, although
this can depend heavily on the objective of the researcher.9 The
first, and oftentimes least desirable prior is noninformed, or a
diffuse prior. This is essentially the assumption that no prior
information is known about the conditional probability of a
phenomena. Thus, any value of the parameter is possible, and
findings are solely modeled on the current data. An example
would be the first space shuttle accident, where there was no
prior information from previous incidents to understand the
probability of such an event occurring. The second type of prior
is the weakly-informed prior which is used when a general ef-
fect direction of the parameter is known, but no specific infor-
mation can be inferred. A semi-informative distribution may
be represented by a standard normal distribution with a large
confidence interval to cover the most likely event. An example
for a weakly informed prior could be nuclear power plant acci-
dents. Very few of these have occurred since the inception of
the nuclear reactor, and reactors vary vastly in their construc-
tion and day-to-day activities, but information can be extracted
from the few incidences that have occurred to understand gen-
eral probabilities. Finally, there is the informed prior, which is
used when specific prior information about a parameter is
available. Our prior distributions are a representative example
of this type of prior, where exact effect sizes and variances are
used based on previously published data. Another example of
this would be with a commercial aviation accident where a da-
tabase of all commercial accidents in modern history can be
accessed for understanding the probabilities of a particular
incident.
Although there is a logical order to the quality of the type of
prior it is important to note that poor application of any type
of prior may lead to a deceptive posterior. However, when ap-
plied correctly, the Bayesian method updates can allow for
previously inferred conclusions with new data to estimate a
more accurate population-level effect. In the case of this study,
we specifically examine whether fatalities occur as a result of a
HEMS accident at a higher rate in concordance with other as-
pects of the flight such as environmental factors or pilot
experience.
Statistical Analysis
Logistic regression is a technique under the generalized lin-
ear model that allows one to model predictions of dichotomous
outcomes with continuous, multinomial, and dichotomous
predictors.19 When the predictors are put into the logistic re-
gression model, they are processed through a logit function that
then compares the effect of their presence on the odds that the
outcome will happen or not. The output of the predictors of the
logistic regression are the log odds. When exponentiated, log
odds become the odds ratio, which is a measurement of an
event occurring vs. an event not occurring based on the out-
come of the dependent variable. The Bayesian framework of
this model is similar in function with the exception that prior
information interacts with the likelihood of each predictor to
produce a posterior distribution of log odds for each variable.
These posterior log odds represent an interval of possible values
that attempts to classify the outcome of the variable.
We conducted a Bayesian logistic regression of whether a
fatality occurred as the result of a HEMS accident predicted by
a set of factors recorded in the associated NTSB reports. Our
analysis utilized prior information of HEMS fatality occurrenc-
es from January 1, 1983, to April 30, 2005, by a previous study.1
Three informative prior coefficients were derived from previ-
ous HEMS fatal accidents. The informative priors were extract-
ed from a logistic regression odds ratio report, then converted
back into logit and standard error coefficients. All binary vari-
ables were standardized with a mean of zero and a maximum of
one as recommended by Gelman et al.’s11 review of Bayesian
logistic regression procedures.
The model was developed using multiple steps including
variable selection, combining prior evidence with our current
data (e.g., likelihood), using a Monte Carlo Markov Chain to
calculate the posterior distribution for each parameter and for
the model outcomes. Variable selection was conducted by ana-
lyzing all variables included in the NTSB reports for which they
were extracted. These variables were then constructed into con-
tingency tables or two-sample comparison tests and analyzed to
determine whether a difference between the two conditions of
the data (i.e., fatality or no fatality) existed. If no difference was
detected, then the variable was left out of the model. We carried
these analyses out using a permutation test, which is a popular
nonparametric testing technique that does not suffer from the
many assumptions (i.e., distribution limitations) that other
nonparametric tests must adhere to. This test randomly resam-
ples the distributions it is examining, while recalculating the
test statistic (e.g., mean, median, etc.) which results in a stron-
ger and more confident test of the differences between the two
samples. This method was conducted due to the low event rate
and irregular distributions (e.g., skewness and kurtosis) of the
data included.
Combining prior and current evidence in the Bayesian
framework often necessitates the use of advanced resampling
methods as simulating and combining distributions can be-
come complex with real-world data. Thus, an extra step in com-
bining the information requires the use of various sampling
techniques such as Monte Carlo Markov Chains (MCMC).
Essentially, this works by estimating samples of the posterior
distributions until an overall distribution is converged upon.
However, as distributions are not always easily estimated, mod-
el diagnostics are available to determine if the sampled posteri-
or distribution is reasonable.
We estimated model fit via the ˆ
R10 (r-hat), the effective sam-
ple size and its ratio,8 and the probability of direction (PoD)14
statistics. The ratios of effective sample size can be interpreted
as the number of independent samples used to calculate
parameters of the posterior distribution over the number of
samples used in the model. ˆ
R is a measurement of MCMC con-
vergence which compares the average variance of each draw of
IP: 192.168.39.211 On: Thu, 16 Sep 2021 16:24:19
Copyright: Aerospace Medical Association
Delivered by Ingenta
BAYESIAN ANALYSIS OF HEMS—Simonson et al.
566 AEROSPACE MEDICINE AND HUMAN PERFORMANCE Vol. 92, No. 7 July 2021
the MCMC to the variance of the pooled draws of all chains. It
measures how consistent each sampled posterior distribution is
to one another. If the models successfully converge to a com-
mon distribution (i.e., the posterior distributions
parameters are similar) then ˆ
R will be 1, indicating successful
posterior draws. Probability of Direction is a calculation that
describes the probability that an effect size will follow a partic-
ular direction. The PoD can be interpreted similarly to a
P-value, where a 97.5 PoD represents a posterior effect in which
97.5% of all values will be in that direction. Finally, we calculat-
ed model comparison of fit via Pareto Smoothed Importance
Sampling-Leave-One-Out (PSIS-LOO) cross-validation to
communicate the strength of evidence about model fit. PSIS-
LOO cross-validation using the LOO package in R20 works by
using the simulated data of the Bayesian model to predict a
single data point, after being trained to predict the rest of the
sample. This method has been shown to be effective in Bayesian
model comparison with a wide array of sample sizes and con-
sistently out-performs other Bayesian model fit methods.21
Like with other model fit procedures (e.g., AIC, BIC, etc.) mod-
el comparison diagnostics are produced for interpretation, in
this case with Expected Log Predictive Density (elpd) and its
standard error. elpd is measured wherein the larger the number
the better the fit, and in measuring the model comparison the
larger the elpd is to its standard error, the better the fit of one
model over the other.
RESULTS
HEMS accidents occurred with an average of 9.29 (N = 131, SD
= 4.46) accidents per year, which is 1.19 more accidents per year
compared to the previous review of HEMS accidents reported
by Baker et al.1 (N = 182, M = 8.1 per year). A one-sample t-test
using the Baker et al.’s1 sample of 8.1 accidents per year as a test
value indicates that significantly more accidents occurred
during the time frame between April 31, 2005, and April 26,
2018, as compared to January 1, 1983, and April 30, 2005.
A total of 398 people were involved in these accidents, 127
(31.90%) were killed and 94 (23.62%) incurred minor, or seri-
ous injuries. The pilots involved in these accidents had accumu-
lated 601,497 (median = 5225, IQR = 3311; 7972) flight hours
in total and 85,681.8 (median = 365.5, IQR = 126.25; 917.75)
flight hours with their respective aircraft at the time of the acci-
dent. Further, the pilots had accumulated 4527 (median = 38,
IQR = 28; 50.75) flight hours in the 90 d before the accident.
Some 73 (77.66%) of the pilots from these accidents held a class
2 medical license and 51 (55.43%) held a medical waiver. There
were 54 (45.38%) accidents which occurred at night, 78 (74.29%)
accidents happened without a patient on board, 15 (15.30%) of
the accidents happened in IFR weather, and 27 (26.21%) result-
ed in a fire. Table I provides a comparison between the demo-
graphic information garnered from our dataset compared to
that of Baker et al.’s.1 Missing data from incomplete or unavail-
able NTSB reports reduced our final sample size to 97 accidents
for our logistic regression model.
Relevant prior information extracted from Baker et al.1 in-
cluded the time of day of the flight, whether the conditions at
the accident site were Instrument Flight Rules (IFR) or Visual
Flight Rules (VFR), and if a post-crash fire occurred. Variable
selection for our logistic regression was determined via multi-
ple criteria. First, any variable that had an informed prior was
included in the model (e.g., time of day, post-crash fire, and
weather conditions). Any other factors extracted from the
NTSB reports for which we could not identify an informed
prior was organized into a contingency table for discrete data
or two-sample comparisons for continuous data and then test-
ed for differences (Table II). The two-sample comparisons of
average pilot flight hours for fatal and nonfatal accidents
(skewness = 1.23; 1.13, kurtosis = 0.54, 0.98) and average pilot
flight hours in the accident helicopter for fatal and nonfatal
accidents (skewness = 2.06; 4.04, kurtosis = 3.40, 17.91) suf-
fered from significant skewness and kurtosis. Thus, each of
these factors were compared via an independent sample per-
mutation test.
In determining our sampling model convergence and fit
(i.e., MCMC), we explore the effective sample size calcula-
tion and ˆ
R calculation listed in Table III. Our Neff/N values
for each parameter estimation of the model is greater than
0.1, indicating a strong effective sample size and confident
estimation of the parameters of our logistic regression mod-
el. Additionally, the ˆ
R of the posterior distributions all con-
verged to 1 indicating that the posterior sample distributions
successfully converged.
Our results show that flying at night (OR 3.06; 95% C.I 2.14,
4.48; PoD 100%), the occurrence of a post-crash fire (OR 18.73;
95% C.I 10.07, 34.12; PoD 100%), and flying under instrument
flight rules (OR 7.54; 95% C.I 3.94, 14.44; PoD 100%) increased
the odds of a crash being fatal. As logistic regression is a classi-
fication technique, we also test how well it classifies the out-
comes compared to the original data as well as compared to the
fitted model with a null model without predictors. In checking
whether the predicted outcomes from the model match the
original data, we can visually check for inconsistencies in a
posterior predictive check. Fig. 2 does not display any apparent
deviance in posterior distributions (y rep) compared to the
distribution of the original data (y). Next, we determine if the
fitted model fits the data better than the null model without any
variables. Essentially, we are testing if the fitted model that deter-
mines the outcome of an accident is doing so better than chance.
A quantifiable means of comparing the models is to measure the
degree of evidence in which the fitted model fits better than the
null model (i.e., using PSIS-LOO). The comparison of fit results
Table I. Current and Prior Accident Demographics.
FAC TO R 1983–2005 2005–2018
Pilot’s Average Flight
Hours
Fatal Non-Fatal Fatal Non-Fatal
5968 6230 6867 5974
Fatal Crashes 71 45
Deaths 184 127
No to Moderate Injuries 373 271
IP: 192.168.39.211 On: Thu, 16 Sep 2021 16:24:19
Copyright: Aerospace Medical Association
Delivered by Ingenta
BAYESIAN ANALYSIS OF HEMS—Simonson et al.
AEROSPACE MEDICINE AND HUMAN PERFORMANCE Vol. 92, No. 7 July 2021 567
for our fitted and null models resulted in an expected log predic-
tive density indicating extreme evidence that the fitted model fits
the data far better than the Null model, elpd = 22.7, SE = 5.4.
In other words, the test shows there is substantial evidence that
the model with the included factors (Table III) predicts fatal out-
comes better than random.
Table III is a summary of the results of the Bayesian logistic
regression model. It includes the odds ratio of the posterior dis-
tribution and the logits and 95% confidence intervals of the
posterior, likelihood, and prior distributions, as well as the
model convergence criteria of the Bayesian logistic regression
model including effective sample size, probability of direction,
and the ˆ
R values.
DISCUSSION
Helicopter Emergency Medical Services (HEMS) expedites the
transport of patients between care facilities or from remote acci-
dent sites to care facilities that may have specialized facilities.
Proponents of HEMS cite the golden hour as evidence favoring
the use of these services.16 The golden hour refers to the finding
that survival rates are highest if a patient receives care as soon as
possible following a severe injury. However, researchers investi-
gating the cost and benefit of HEMS transport have argued that
these services are being utilized for patients with conditions that
do not justify the cost or potential risk of this mode of transport.
For instance, a meta-analysis showed that 61.4% of the patients
who were transported via helicopter from a scene to a hospital
only received minor injuries as a result of the original incident.5
Additionally, the trends in HEMS accidents should be cause
for concern given the risk to passengers, aircrew, medical crew,
Table II. Variable Comparison with Permutated P-values.
FAC TO R NOT FATAL FATAL P-VALUE (95% C.I.)
Passenger on Board
Yes 6 3 P = 0.733 (0.721, 0.745)
No 57 38
Medical Class
2 42 31 P = 0.609 (0.596, 0.622)
1 14 7
Patient on Board
Yes 13 14 P = 0.174 (0.164, 0.184)
No 50 28
Medical Waiver
Yes 29 22 P = 0.830 (0.820, 0.839)
No 25 16
NVG’s Used
Yes 5 3 P = 0.077 (0.079, 0.085)
No 3 11
Avg. Pilot Total Flight Hours median = 5380, Q1 = 3380;Q3 = 7928 median = 5092, Q1 = 3200;Q3 = 9260 P = 0.295 (0.283, 0.307)
Avg. Pilot Aircraft Flight Hours median = 480, Q1 = 125;Q3 = 760 median = 315, Q1 = 130;Q3 = 1100 P = 0.11 (0.102, 0.118)
Table III. Final Model.
FAC TO R
ODDS
RATIO
POSTERIOR
(LOGIT; 95% C.I.)
LIKELIHOOD
(LOGIT; 95% C.I.)
PRIOR
(LOGIT, SE)
EFF.
SAMPLE SIZE PoD
ˆ
R
Intercept −2.15; −2.78, −1.58 −2.13; −3.25, −1.24 Cauchy (0,10) 4691 100 1
Night Flight 3.06 1.12; 0.76, 1.50 0.99; −0.15, 2.24 Normal (1.16, 0.211) 4192 100 1
Fire 18.73 2.93; 2.31, 3.53 3.39; 2.08, 5.02 Normal (2.77, 0.355) 4785 100 1
IFR Weather 7.54 2.03; 0.1.37, 2.67 1.86; 0.35, 3.57 Normal (2.079, 0.377) 5037 100 1
Fig. 2. Fitted model graphical posterior predictive density
overlay.
IP: 192.168.39.211 On: Thu, 16 Sep 2021 16:24:19
Copyright: Aerospace Medical Association
Delivered by Ingenta
BAYESIAN ANALYSIS OF HEMS—Simonson et al.
568 AEROSPACE MEDICINE AND HUMAN PERFORMANCE Vol. 92, No. 7 July 2021
and the patient as HEMS accidents have historically suffered
from the highest fatality rate of any other aviation transporta-
tion method.6,12 Our model offers insight into the factors that
may influence increased fatalities given that an accident has oc-
curred. The results show that a post-crash fire, flying in IFR
conditions, and flying at night are more frequently associated
with accidents involving fatalities. Further, due to the applica-
tion of Bayesian inference, we were able to combine the mea-
surable influence of information from prior analyses. The prior
influences of post-crash fires, night flying, and flying in poor
weather strengthened our inferences of their effects on the
chance of a fatal accident due to the information they added to
our model. Thus, as a result of our Bayesian specified model, we
present a review analysis of the odds of these factors occurring
during fatal accidents from January 1st, 1983, to April 26, 2018,
which is the most extensive review of HEMS accidents to date.
Additionally, inspection of the prior and current evidence of
the included factors indicates that over these past 35 yr of
HEMS accidents, the odds that a factor will increase the likeli-
hood of a fatality as a result of an accident has barely changed
(Table III). This indicates that the interventions and policies
implemented in the United States, such as those suggested by
the 2009 NTSB HEMS review forum,6 may not have resulted in
sufficient safety improvements in relation to the identified fac-
tors. However, this review is limited only to the prior informa-
tion available to us and the data gathered from NTSB accident
reports.
These results advance our understanding of the factors asso-
ciated with the outcomes of HEMS flights. While the estimated
rate of HEMS accidents is seemingly low given the number of
flights per year, the risk of death, or injury, as a result of any of
these accidents is high. Specifically, the environmental condi-
tions included in the analyses (flying at night and in IFR weath-
er) are shown to increase the risk of a fatal accident in both gen-
eral aviation and HEMS but are also shown with high confidence
to contribute to higher rates of fatalities.1,2 One caveat is that the
temporal order of these factors is not considered in this analysis.
Specifically, flying in IMC weather or at night precede a post-
crash fire, and thus may be considered causal factors. However,
the analysis structure (e.g., choice and implementation of prior
information) limited our ability to conduct such analyses.
The analysis has several limitations pertaining mainly to the
lack of prior information. With respect to the priors, the field of
HEMS accident investigation lacks data on a number of import-
ant issues including the identification of the factors associated
with an increased number of fatalities as a result of an accident,
factors associated with increased numbers of injuries as a result
of an accident, and most importantly a comparison between fac-
tors associated with accidents and nonaccidents. The absence of
flight data activity (i.e., yearly total of HEMS flights, environ-
mental conditions on all HEMS flights) regardless of the
outcome prevents us from understanding if the conditions pres-
ent during a HEMS accident are unique to those incidents.
Future research should explore the increased risk HEMS flights
pose to patients, passengers, and care providers.
We presented a comprehensive analysis of what factors are
associated with fatal HEMS accidents by combining prior evi-
dence with current information to form a confidence interval of
possible odds ratios associated with each factor. We found that
over the past 35 yr post-crash fires, flying at night, and flying in
inclement weather conditions were consistently associated with
a higher likelihood of a fatality occurring. Further, we argue that
the lack of data about various aspects of the HEMS community
prevent us from understanding the risk a HEMS transport plac-
es on the patients, providers, and crew. Increasing patient safety
during HEMS flights and ensuring provider and crew safety
during flights is of paramount importance. A framework of re-
search investigating the data-driven analysis of accident risk
and added risk to patients should be the focus of future work.
ACKNOWLEDGMENTS
The authors would like to acknowledge: Loren Groff, Ph.D., of the National
Transportation Safety Board (NTSB) for his insights and supporting role re-
garding the use of the NTSB accident database; Eileen Frazer, executive director
of the Commission on Accreditations of Medical Transport Systems, for her
assistance and supporting role on data collection; and Levi Demaris of Emb-
ry-Riddle Aeronautical University for his assistance and supporting role in data
collection and organization.
Financial Disclosure Statement: The authors have no competing interests to
declare.
Authors and Affiliations: Richard J. Simonson, B.S., Joseph R. Keebler, Ph.D.,
and Alex Chaparro, Ph.D., Embry-Riddle Aeronautical University, Daytona
Beach, FL.
REFERENCES
1. Baker SP, Grabowski JG, Dodd RS, Shanahan DF, Lamb MW, Li GH.
EMS helicopter crashes: what influences fatal outcome? Ann Emerg Med.
2006; 47(4):351–356.
2. Bensyl DM. Factors associated with pilot fatality in work-related aircraft
crashes, Alaska, 1990-1999. Am J Epidemiol. 2001; 154(11):1037–1042.
3. Bledsoe BE. Air medical helicopter accidents in the United States: a five-
year review. Prehosp Emerg Care. 2003; 7(1):94–98.
4. Bledsoe BE, Smith MG. Medical helicopter accidents in the United States:
a 10-year review. J Trauma. 2004; 56(6):1325–1328.
5. Bledsoe BE, Wesley AK, Eckstein MM, Dunn TF, O’Keefe M. Helicopter
scene transport of trauma patients with nonlife-threatening injuries: a
meta-analysis. J Trauma. 2006; 60(6):1257–1265.
6. Blumen IRA. An analysis of HEMS accidents and accident rates. [Inter-
net] 2009 [Accessed 2019 September 26]. Available from: https://www.
ntsb.gov/news/events/Documents/NTSB-2009-8a-Blumen-revised-fi-
nal-version.pdf.
7. Federal Aviation Administration. General aviation and part 135 activity
surveys. 2017. [Accessed 2019 September 26]. Available from https://
www.faa.gov/data_research/aviation_data_statistics/general_aviation/.
8. Gabry J, Modrák M. Visual MCMC diagnostics using the Bayesplot
package. Bayesplot. Published 2020. [Accessed September 27, 2020.]
Available from: https://mc-stan.org/bayesplot/articles/visual-mcmc-
diagnostics.html.
9. Gelman A. Prior distribution. In: El‐Shaarawi AH, Piegorsch WW,
Hallin M, editors. Encyclopedia of Environmetrics. Hoboken (NJ):
John Wiley & Sons, Ltd; 2006.
IP: 192.168.39.211 On: Thu, 16 Sep 2021 16:24:19
Copyright: Aerospace Medical Association
Delivered by Ingenta
BAYESIAN ANALYSIS OF HEMS—Simonson et al.
AEROSPACE MEDICINE AND HUMAN PERFORMANCE Vol. 92, No. 7 July 2021 569
10. Gelman A, Carlin JB, Stern HS, Dunson DB, Vehtari A, Rubin DB. Basics
of Markov chain simulation. In: Bayesian Data Analysis. Boca Raton (FL):
CRC Press; 2013:285.
11. Gelman A, Jakulin A, Pittau MG, Su Y-S. A weakly informative default
prior distribution for logistic and other regression models. Ann Appl Stat.
2008; 2(4):1360–1383.
12. International Helicopter Safety Team. 2017. [Accessed 2019 September
26]. Available from: http://www.ihst.org/portals/54/symposium/2017%20
IHST%20Regional%20Partners%20Session130.pdf.
13. Lerner EB, Moscati RM. The golden hour: scientific fact or medical “ur-
ban legend”? Acad Emerg Med. 2001; 8(7):758–760.
14. Makowski D, Ben-Shachar MS, Chen SHA, Lüdecke D. Indices of effect
existence and significance in the Bayesian framework. Front Psychol.
2019; 10:2767.
15. Miranda AT. Understanding human error in naval aviation mishaps. Hum
Factors. 2018; 60(6):763–777.
16. Pham H, Puckett Y, Dissanaike S. Faster on-scene times associated with
decreased mortality in helicopter emergency medical services (HEMS)
transported trauma patients. Trauma Surg Acute Care Open. 2017;
2(1):e000122.
17. Reddick EJ. Evaluation of the helicopter in aeromedical transfers. Aviat
Space Environ Med. 1979; 50(2):168–170.
18. Solomon SS, King JG. Influence of color on fire vehicle accidents. J Safety
Res. 1995; 26(1):41–48.
19. Tabachnick BG, Fidell LS, Ullman JB. Using multivariate statistics. New
York (NY): Pearson; 2019.
20. Vehtari A, Gabry J, Yao Y, Gelman A. The R project for statistical comput-
ing. The R Foundation; 2019. Available from: https://cran.r-project.org/
web/packages/loo/index.html.
21. Vehtari A, Simpson D, Gelman A, Yao Y, Gabry J. Pareto smoothed im-
portance sampling. arXiv. 2019; available from: http://arxiv.org/abs/
1507.02646.
