applying artificial neural networks to predict ... · original research: empirical research –...
TRANSCRIPT
ORIGINAL RESEARCH: EMPIRICAL RESEARCH –QUANTITATIVE
Applying artificial neural networks to predict communication risks in
the emergency department
Annamaria Bagnasco, Anna Siri, Giuseppe Aleo, Gennaro Rocco & Loredana Sasso
Accepted for publication 13 April 2015
Correspondence to A. Bagnasco:
e-mail: [email protected]
Annamaria Bagnasco MSC PhD
Nursing Researcher
Department of Health Sciences, University
of Genoa, Italy
Anna Siri PhD
Researcher
School of Medical and Pharmaceutical
Sciences, University of Genoa, Italy
Giuseppe Aleo MA
PhD Student in Epidemiology
Department of Health Sciences, University
of Genoa, Italy
Gennaro Rocco MSC RN NP
Director of the Centre of Excellence for
Nursing Scholarship
Centre of Excellence for Nursing
Scholarship, Rome, Italy
Loredana Sasso MA MSC RN
Associate Professor of Nursing
Department of Health Sciences, University
of Genoa, Italy
BAGNASCO A . , S IR I A . , ALEO G . , ROCCO G . & SAS SO L . ( 2 0 1 5 ) Applying arti-
ficial neural networks to predict communication risks in the emergency depart-
ment. Journal of Advanced Nursing 00(0), 000–000. doi: 10.1111/jan.12691
AbstractAims. To describe the utility of artificial neural networks in predicting
communication risks.
Background. In health care, effective communication reduces the risk of error.
Therefore, it is important to identify the predictive factors of effective
communication. Non-technical skills are needed to achieve effective
communication. This study explores how artificial neural networks can be applied
to predict the risk of communication failures in emergency departments.
Design. A multicentre observational study.
Methods. Data were collected between March–May 2011 by observing the
communication interactions of 840 nurses with their patients during their routine
activities in emergency departments. The tools used for our observation were a
questionnaire to collect personal and descriptive data, level of training and experience
and Guilbert’s observation grid, applying the Situation-Background-Assessment-
Recommendation technique to communication in emergency departments.
Results. A total of 840 observations were made on the nurses working in the
emergency departments. Based on Guilbert’s observation grid, the output
variables is likely to influence the risk of communication failure were
‘terminology’; ‘listening’; ‘attention’ and ‘clarity’, whereas nurses’ personal
characteristics were used as input variables in the artificial neural network model.
A model based on the multilayer perceptron topology was developed and trained.
The receiver operator characteristic analysis confirmed that the artificial neural
network model correctly predicted the performance of more than 80% of the
communication failures.
Conclusion. The application of the artificial neural network model could offer a
valid tool to forecast and prevent harmful communication errors in the
emergency department.
Keywords: artificial neural network, communication skills, emergency depart-
ment, nursing, prediction, risk
© 2015 John Wiley & Sons Ltd 1
Introduction
Several studies on communication relations in Emergency
Departments report that approximately 89% of the profes-
sional activities require good communication skills (Spencer et al.
2002) and that the use of standardized communication tech-
niques, such as the Situation-Background-Assessment-Recom-
mendation (SBAR) technique (Haig et al. 2006), in clinical
situations requiring immediate interventions, is mandatory for
patient safety (Meginniss et al. 2012). Clear and complete com-
munication and the constant check that communication is prop-
erly understood are instrumental to ensure patient safety and
the quality of care (Spencer et al. 2002, Jones et al. 2013).
A study, previously conducted by our research group, per-
mitted to identify and categorize the communication failures
detected in the nurse–patient communication interactions in
the Emergency Department and their coding according to the
Joint Commission International (JCI) standards (Bagnasco
et al. 2013). We used the failure mode and effects analysis
(FMEA), because it effectively assesses and prioritizes the
areas of risk in clinical practice (Day et al. 2006, Lago et al.
2012). The preliminary phase of that study, after a review of
the literature, led to the identification of an observation grid
on ‘Communication Performance’ by Jean-Jacques Guilbert
(Guilbert 1990, 2002) (Figure 1), a validated tool used for
the observation of health workers’ communication perfor-
mance based on ‘Terminology’ (i.e. whether words used were
precise and adequate), ‘Listening’ (i.e. listening and correctly
understanding a message), ‘Attention’ (i.e. paying verbal and
behavioural attention by providing feedback) and Clarity
(i.e. communicating in a clear and precise manner so that
information is correctly understood. With this observation
grid, we observed the nursing staff working in the Emergency
Departments and measured the quali-quantitative require-
ments that needed to be implemented to improve their non-
technical skills (Flin et al. 2003) and in particular internal
and external communication (Kilner & Sheppard 2010).
‘Communication Skills Training’ can improve emergency
nurses’ communication and empathy skills associated with
an increase inpatient satisfaction and a reduction in unde-
sirable events and complaints during nurse–patient interac-
tions (Ak et al. 2011). Therefore, in 2013 our team
conducted a study on communication skills training in an
advanced simulation centre (Bagnasco et al. 2014) to cor-
rect communication failures. In this case Guilbert’s observa-
tion grid was used for educational purposes.
Instead, in this study, Guilbert’s observation grid (Fig-
ure 1) was used with the artificial neural networks (ANNs)
to predict nurse–patient communication risks from data col-
lected from three Adult Emergency Departments in Central
Italy between March 2011–May 2011.
What are artificial neural networks?
ANNs are mathematical models that try to imitate the func-
tional processes of the human brain. ANNs involve non-
linear relationships among different data sets that cannot
always be fully identified through the use of conventional
linear analyses. ANNs are extensively used in technology
and science with applications in chemistry, physics, biology
and medicine. As reported by Amato et al. (2013), ANNs
manage several types of input data to produce a clinically
relevant output, for example, the probability of a certain
pathology or classification of biomedical objects. These data
are processed in the ‘setting’ of previous training history on
a defined sample database.
ANNs are known to be a potent tool to simulate several
non-linear systems and have been applied to numerous
Why is this research or review needed?
� Communication failures in the emergency department con-
tinue to raise serious safety issues. Artificial neural net-
works can be successfully applied to predict and avoid
communication failures.
� Artificial neural networks are best suited for cases, where
there is a known relationship between inputs and outputs,
but their precise nature is not clear.
What are the key findings?
� Artificial neural networks can extract patterns and trends
from non-linear data sets that are too difficult for conven-
tional computers to calculate.
� Artificial neural networks are able to include uncertainty
in data sets.
� Artificial neural networks can be ‘trained’ using fewer
points than any other statistical model.
How should the findings be used to influence policy/practice/research/education?
� Artificial neural networks are demonstrated to be an effec-
tive tool for predicting the risk of communication failures,
but further research is needed.
� Artificial neural networks can be used in general as a risk
prediction tool.
� Early prediction of risk enables to arrange targeted educa-
tional interventions before harmful situations actually
occur.
2 © 2015 John Wiley & Sons Ltd
A. Bagnasco et al.
problems of statistically significant complexity in many
fields, including engineering, agriculture (Rodriguez Galdon
et al. 2010), education (Siri 2014), medicinal chemistry
(Pandini et al. 2013), diagnostics (Qeethara Kadhim 2011,
Irfan Khan et al. 2013) and pharmaceutical research (Weso-
lowski & Suchacz 2012). In medicine, ANNs have been
successfully applied to various areas, such as diagnostic sys-
tems (Dey et al. 2011, Hu et al. 2013, Liu & Jiang 2013),
biochemical analysis (Catalogna et al. 2012, Fernandez de
Canete et al. 2012), image analysis (Barbosa et al. 2012,
Saghiri et al. 2012) and drug development (Li et al. 2005,
Patel et al. 2012). In diagnostic systems, ANNs are usually
exploited to identify cancer and heart diseases. In biochemi-
cal analysis, ANNs have been used to examine glucose lev-
els in people affected by pathological conditions such as
tuberculosis and diabetes or in the analysis of medical
images.
In the emergency department, ANNs have predominantly
been used to quantitatively forecast patient arrivals and
admission rates (Champion et al. 2007, Jones et al. 2008,
Schweigler et al. 2009, Xu et al. 2013), or to obtain feed-
back to support clinical decisions (Hollander et al. 2004).
To our knowledge, ANNs have never been used to predict
and therefore prevent potentially harmful nurse–patient
communication interactions to increase patient safety,
which is the purpose of the present article.
Why use artificial neural networks?
ANNs are best suited for cases, where there is a known
relationship between the variable inputs and outputs, but
the particular nature of the connection is not clear. ANNs
are helpful when the link between several different variables
needs an accurate mathematical model that still has not
been developed.
The added value of ANNs is that they extract patterns
and trends from data sets that are too complex to build
with conventional computers. Another advantage of ANNs
is their ability to include uncertainty or ‘noise’ in a given
data set. Furthermore, ANNs make no assumptions about
the statistical nature of the data and can integrate nominal
and ordinal data. ANNs can be trained using comparatively
fewer points than any other statistical model and it is not
necessary to select a data distribution model.
What do we mean by ‘training’ a Neural Network?
Biologically, neural networks are made up of large numbers
of neuron cells deeply interconnected with one another. In
fact, each neuron is a specialized component that sends out
specific electrochemical signals and the axons of each neu-
ron are linked to the dendrites of other neurons through
synapses. The signals are sent through this intricate mecha-
Terminology Listening Attention Clarity
–2 Too detailed – not
adequate
Hears but not listens Not pay attention
neither verbally nor
through behaviour
Communication not
clear and information
not precise
–1
0
+1
Too detailed –
adequate
Listens but not repeats Verbal attention is
inconsistent with
behavioural one
Communication not
clear and information
is partially precise
Adequate – Not precise Listens but not always
repeats correctly
Pays verbal attention Communication not
always clear and
information not always
precise
Precise terminology
but not replies
immediately to
questions
Listens and repeats
correctly
Pays verbal and
behavioural attention
Communication not
always clear but
information is precise
+2 Replies immediately to
questions
Checks understanding Feedback to
interlocutor
Information is correctly
understood
Figure 1 Guilbert’s observation grid on communication performance. The scores in first column on the left measure the level of risk.
Source: J.-J. Guilbert. Guide p�edagogique pour les personnels de sant�e. Geneva, World Health Organization, 1990.
© 2015 John Wiley & Sons Ltd 3
JAN: ORIGINAL RESEARCH: EMPIRICAL RESEARCH – QUANTITATIVE Risk prediction in ED using ANNs
nism, which is fundamentally regulated by synapses. In his
study, Hebb (1949) showed that the ‘learning process’
mainly consists in changing the strength of several synaptic
connections. The signals that pass produce a response and
the synapses carefully regulate the way it is transferred.
Artificial neural networks (ANNs) try to imitate the
behaviour of the brain. Similar to the biological nervous
system, also ANNs contain layers of simple processors (or
nodes) of data that interact through weighted connection
lines and generate an outcome. The weight-balance of these
lines is done through a ‘training session’ of input data that
the network uses to adapt its links. In short, ANNs during
training attempt to extract patterns, trends and relation-
ships from data sets starting from various input and output
variables. Obviously, the structure of an ANN is much sim-
pler than that of a biological nervous network, with many
fewer nodes or interconnections.
Researchers have explored artificial intelligence over the
last 70 years. McCulloch and Pitts carried out the first stud-
ies on ANNs in 1943. In 1959, Rosenblatt formulated the
first learning algorithm, known as the perceptron that was
only a solution to simple linear problems. In 1974, Werbos
reported the first non-linear processing capabilities of
ANNs. Thanks to the growing progresses in computer tech-
nology, the scientific community discovered the back propa-
gation algorithm – which is a network training or learning
algorithm – and its advantages in computational power.
Background
Our research team started investigating the important role
of communication skills for ensuring safety in the Paediatric
Emergency Department setting in 2008 (Bagnasco et al.
2013). During that study, many nurses were found to have
poor communication skills, therefore, our research team
conducted another study to improve communication skills
through educational interventions in an Advanced Simula-
tion Centre (Bagnasco et al. 2014).
However, educational interventions on their own are not
enough to ensure safe communication in the Emergency
Department, but needed to find a tool that could predict
communications risks before they actually occurred. There-
fore, ANNs were identified as a possible tool that could
reliably achieve this purpose.
The Artificial Neuron
As described by Bardan (2014), ANNs are made up of neu-
rons and each neuron consists of a ‘summation block’ and an
‘activating block’. The input variables are multiplied by the
optimized weight parameters (w) and their products are
summed. The neuron output is attained through the transfor-
mation of the sum using an activation function. Perceptron
networks commonly use continuous sigmoidal activation
functions that convert the output value into a range between
0-1. From this perspective, the artificial neuron is similar to
regression models. In fact, a bias weight (‘constant’ in statisti-
cal terminology) is used to optimize the final prediction, simi-
lar to regression models. However, regression models have
an important weakness that consists in the inability to reveal
intricate relationships in given data sets. Moreover, the corre-
lation type between inputs and outputs has to be chosen
before starting the analysis (Bardan 2014) (Figure 2).
In 2000, Zhang and Gupta described that the most com-
mon network structure is the multilayer perceptron (MLP),
where the neurons are grouped into layers. No layers
Inputs
Threshold (bias)
Weights
Sum-of-Product
Outputγ
χ
Θ
∑ ƒ(χ)
Summationfunction
Activationfunction
A1
W1
W2
W3
WN
1
A2
A3
AN
Figure 2 Visual display of an artificial neuron. The inputs are multiplied with the optimized weights and thereafter summed. The sum is
transformed by the neuron’s activation function into the output value. Source: Own elaboration.
4 © 2015 John Wiley & Sons Ltd
A. Bagnasco et al.
(levels) of neurons communicate with each other. The first
and last layers are respectively named ‘Input’ and ‘output’,
because they represent inputs and outputs of the whole net-
work. The ‘Hidden layers’ are the remaining layers that are
situated inside the network, between the input and the out-
put layers. Usually, an MLP neural network contains an
input layer, one or more hidden layers and an output layer
(Figure 3).
Statistically significant parameters for an accurate predic-
tion are the number of hidden layers and the number of neu-
rons on each level. The output layer generates the network
response that commonly consists of one or two neurons.
The study
This study consists of four phases. In Phase 1, we collected
the data through the observations of the nurse–patient com-
munication interactions and measured communication risk
with Guilbert’s observation grid (Figure 1). In Phase 2, we
trained the ANNs using the input (Table 1) and output
variables (Table 2). In Phase 3, we tested the predictive per-
formance of the ANNs (Figure 5). In Phase 4, we used the
ROC analysis (Figure 4) to understand whether the predic-
tive performance of the ANNs was correct.
Aim
The aim of this paper was to study the usefulness of ANNs
in predicting the risk of communication failures in EDs
through selected variables.
Rationale
Based on our previous studies (Bagnasco et al. 2013, 2014),
two tools were identified to collect data: (1) A question-
naire to collect nurses’ personal characteristics; (2) the Situ-
ation-Background-Assessment-Recommendation (SBAR)
technique applied to communication in the Emergency
Department, which investigated the communication behav-
iours of the nurses with patients (and their family members)
using the four dimensions of Guilbert’s observation grid
(i.e. terminology, listening, attention and clarity).
Therefore, to test ANNs as a risk prediction tool for
communication failures in the ED we asked two research
questions, one for each tool:
1 How accurately do nurses’ personal characteristics predict
the risk of communication failures?
2 Which of the four communication dimensions weigh most
in predicting the risk of communication failures?
Design
A multicentre observational study.
Sample
We selected a convenience sample of 840 nurses working
in the emergency departments (EDs) from three different
hospitals in Central Italy. After contacting the Nursing
Inputlayer
Hiddenlayer
Outputlayer
Synapses
Figure 3 The layers of an artificial neural network. Source: Own
elaboration.
Table 1 Input variables and respective description. The nominal
variables are qualitative characteristics that cannot be numerically
ordered. Instead, ordinal variable are characteristics that can be
numerically ordered, such as level of education and school marks.
Input variables Description of the input variables Type
M/F Gender Nominal
Age Age bracket Ordinal
Seniority Years of service in current Emergency
Department
Ordinal
Professional
Qualification
Title of professional qualification Nominal
Postgraduate
qualification
Titles of Postgraduate qualification
(masters, specialization, etc.)
Nominal
Structure Hospital where nurses worked Nominal
Source: Own elaboration.
Table 2 Output variables.
Output variables Type
Terminology Ordinal
Listening Ordinal
Attention Ordinal
Clarity Ordinal
Source: Own elaboration.
© 2015 John Wiley & Sons Ltd 5
JAN: ORIGINAL RESEARCH: EMPIRICAL RESEARCH – QUANTITATIVE Risk prediction in ED using ANNs
Directors of the three hospitals and obtaining their
approval for conducting this observational study involving
the nurse–patient communication interactions, our research
team illustrated the project to the head nurse of each emer-
gency department (ED). The head nurses then asked if any
nurses working in the ED on a full time basis volunteered
to take part in this study. The members of our research
team provided preliminary training both to each head nurse
and to their group of nurses who had volunteered to take
part in this study. Preliminary training was conducted in
small groups and consisted of a four-hour theoretical and
practical training session to make sure that the tools (the
personal characteristics questionnaire and Guilbert’s obser-
vation grid) were used correctly and that observations were
conducted in a rigorous and objective manner. This preli-
minary training session was essential both to produce com-
parable data and ensure good inter-rater reliability.
Data collection
A multicentre observational study database was used for this
research. Data were collected by observing nurse–patient
1·0Terminology Listening
Attention Clarity
·8
·6
·4
·2
·2 ·4 ·6 ·8 1·01 - Specificity
Area under curve1 ·910
·917·927
23
Area under curve1 ·949
·915·952
23
Area under curve1 ·910
·866·915
23
Area under curve1 ·911
·886·924
23
Sen
sitiv
ity
·0·0
1·0
·8
·6
·4
·2
·2 ·4 ·6 ·8 1·01 - Specificity
Sen
sitiv
ity
·0·0
1·0
·8
·6
·4
·2
·2 ·4 ·6 ·8 1·01 - Specificity
Sen
sitiv
ity
·0·0
1·0
·8
·6
·4
·2
·2 ·4 ·6 ·8 1·01 - Specificity
Sen
sitiv
ity
·0·0
123
123
123
123
Figure 4 ROC curve of the four dimensions of communication. Source: Own elaboration.
6 © 2015 John Wiley & Sons Ltd
A. Bagnasco et al.
communication interactions in 840 nurses while carrying
out their routine activities in the Emergency Department.
Observations were made over a period of 3 months
(March–May 2011), until at least 200 communication inter-
actions were collected in each hospital. Therefore, the
groups of appropriately trained nurses observed the com-
munication behaviours of their nurse colleagues with
patients during their routine work activities in the ED. This
strategy to train nurses already working in the ED instead
of having a person from outside the hospital (i.e. a member
of our research team), to minimize the Hawthorne effect
(Landsberger 1958).
The tool used for observation consisted of two parts:
1 A questionnaire to collect data on the facility, field of
investigation and characteristics of the nurses included in
the sample.
2 Guilbert’s observation grid (Guilbert 1990) based on the
four communication dimensions.
The purpose of the first part of the tool was to collect the
ED staff nurses’ personal and descriptive data and their levels
of training and experience (i.e. age, gender, years of service in
the current workplace, professional qualification, postgradu-
ate education and type of structure). The second part of the
tool investigated the communication behaviours of the ED
staff nurses with patients in relation to four dimensions:
terminology, listening, attention and clarity (Figure 1). For
each of these dimensions, a rating scale (ranging from �2-+2)
measured the level of communication risk.
Ethical considerations
This study received Research Ethics Committee approval
from the University of Genoa. All participants took part
on a voluntary basis and were appropriately informed
before being included in the study. In addition, all partici-
pants spontaneously accepted to reciprocally observe one
another.
Data analysis
The database of 840 nurses working in three Accident and
Emergency Departments in Rome consisted of a total of 10
variables: the four dimensions of the observation grid (i.e.
terminology, listening, attention and clarity; these were the
output variables) and six regarding age, gender, working
seniority, level of education, professional qualifications and
the structure (place), where they worked (these were the
input variables). Not all the variables included in the model
were suitable at the first attempt to train the ANN.
The six main input variables included in this study are
shown in Table 1. The output variables were the four
dimensions included in Guilbert’s observation grid related
to communication in the ED: terminology, listening, atten-
tion and clarity (Table 2). The output layer consisted of a
variable that indicated communication risk, which in the
ANNs was summarized into three units (1 = real risk;
2 = moderate risk; 3 = no risk).
To increase ANN predictive performance, a correlation
analyses was conducted between all the variables to identify
any strong correlations between the different variables
included and remove redundant information from the data-
base. This analysis showed that there were strong obvious
correlations with age and seniority and weak correlations
with gender, education, professional qualifications and the
structure where they worked. Microsoft Office Excel, Ver-
sion 2007 and IBM SPSS Statistics for Windows, Version
22�0 (IBM Corp. Released 2013, Armonk, NY, USA) were
used for data processing and running the ANNs.
ANN Topology selection and training
The identification of an appropriate topology (i.e. the num-
ber of hidden layers and the number of nodes per hidden
layer) is one of the most crucial aspects of the application
of neural networks. The neural network architecture used
in this study is the multi-layered feed-forward network
model that is widely and successfully applied for classifica-
tion, forecasting and problem-solving (Irfan Khan et al. 2013).
In general, a network topology is often decided only after
a lengthy process of trial-and-error. The approach we
adopted considered the training of many different networks
and the subsequent selection of the ‘best’ one. All neural
networks are trained until the error (sum of squared error)
for each training session reached the minimum, before it
started to rise again.
After topology determination, the data were normalized
finding a mean and a standard deviation of each feature
and for every cell of each row to subtract its mean and
divide it by the standard deviation of the right feature. The
resulting data have a zero mean and a standard deviation
equal to one.
Therefore, the data are separated into training and vali-
dation subsets. The ANN is trained with a standard back-
propagation learning algorithm. The activation function
used for each node in the network is the binary sigmoidal
function demarcated (with r = 1) as output = 1/(1 + e�x),
where x is the sum of the weighted inputs to that specific
node. This function restricts the output of all nodes in the
network to the range 0-1. Between each training epoch, the
error rate is computed for data validation.
© 2015 John Wiley & Sons Ltd 7
JAN: ORIGINAL RESEARCH: EMPIRICAL RESEARCH – QUANTITATIVE Risk prediction in ED using ANNs
Due to the risk of overfitting the neural network too eas-
ily, producing an error rate on validation data larger than
the error rate on the training data, it is crucial not to over-
train the data. Overtraining is when you train the network
for so long that it starts to memorize the training set. The
validation error reduces in the early epochs but after some
time, it begins to rise. The point of minimum validation
error is a valid indicator of the best number of epochs for
training. Our results indicated that the minimum error of
the validation set could be obtained in approximately 500
epochs. The training results are certainly dependent on the
initial values of the weights that are generally random.
We examined the performance of the test across the
entire range of possible values using the curves of the recei-
ver operator characteristic (ROC) analysis. The ROC curve
is a plot of the true positive rate (sensitivity) vs. the false
positive rate (1-specificity). A perfect test would show
points in the upper-left-corner, with 100% sensitivity and
100% specificity (Woods & Bower 1997). An area under a
curve (AUC) of 1 means that the test is perfect, whereas an
AUC of 0�5 means that the test has a probability of 50% to
classify (similar to the toss of a coin). A diagnostic test is
considered adequate with an area under the curve >80%.
Validity and reliability
The validity and reliability of the data sets collected during
the communication observations made at the beginning of
this was ensured through the observers’ use of Guilbert’s
observation grid, by attributing a score between �2-+2 to
the four dimensions of communication: terminology; listen-
ing; attention and clarity (Figure 1). For this purpose, a 4-
hour preliminary training session was provided to the ED
nurses who were going to observe the nurse–patient com-
munication interactions of their peers.
Results
A total of 840 observations were made on the nurses work-
ing in the EDs, with a good distribution between males and
females. Most of the observations involved nurses aged
between 31-40 years. With regard to seniority (years of ser-
vice in the current ED) most of the nurses had worked
between 6-10 years (Table 3). The other three variables
(education, professional qualification and structure) are not
included because we found that they did not have an
impact on the prediction of communication risk.
The ANN’s learning ability was checked through the cal-
culation of the mean-squared error, as error measure. We
used a learning algorithm called ‘Back Propagation’, which
modified the synaptic weights to minimize global error
made by the ANN, measured with the mean squared error
(MSE) between the output of the network and the correct
results. The network was trained on 80% of the population
included in the study; 20% of the population was used as a
pilot test.
The network considered to have the highest level of pre-
diction is the multi-perceptron network with 35 units, one
hidden layer and 10 units in the hidden layer. The four
dimensions of communication (clarity, attention, listening
and terminology) were analysed separately to determine the
importance of the independent variables for each one and
compare them with one another. We examined the perfor-
mance of the test across the entire range of possible values
using the ROC curve analysis (Figure 3). Regarding the
clarity of communication, training was performed on 684
cases, causally chosen (81�4% of the cases) and the test on
the remaining 156 (18�6%).
In our case, the values using the ROC curve analysis
were higher than 90%. Therefore, the test was very accu-
rate for the risk levels 1 and 3 and moderately accurate for
risk level 2. During training, the neural network correctly
predicted 82�5% of the cases and 86�5% during the test.
Concerning the attention to communication, the training
was performed on 669 cases, chosen on a casual basis
(79�6% of the cases) and the test was conducted on the
remaining 20%, equal to 171 (20�4%). The values using
the ROC curve analysis were >90% and, therefore, corre-
sponded to a probability that was 90% higher than a cor-
rect classification. Therefore, the test was very accurate for
the risk levels 1 and 3 and moderately accurate for risk
level 2. The neural network predicted correctly during
training in 81% of the cases and 83�6% during the test.
Pertaining to the terminology dimension, the training was
performed on 689 cases, causally chosen (82% of the cases)
Table 3 The most significant results of the observations made on
the communication interactions of 840 nurses in three different
Emergency Departments with their patients between March–May
2011.
Gender Age
Seniority
(years of service)
Males = 40�6%Females = 59�4%
31-40 years = 41�3%41-50 years = 32�9%51-60 years = 7�5%>61 years = 0�8%
0-5 years = 26�2%6-10 years = 37%
11-15 years = 22%
16-20 years = 8�2%>20 years = 6�05%
The other three variables (education, professional qualifications
and place of work) were excluded because they were found to have
no impact on the prediction of communication risk.
8 © 2015 John Wiley & Sons Ltd
A. Bagnasco et al.
and the test on the remaining 151 (18%). Also in this case,
the test was very accurate. The neural network predicted
correctly during training in 81�9% of the cases and 81�5%during the test.
Concerning the listening dimension, the training was per-
formed on 675 cases, causally chosen (80�4% of the cases)
and the test on the remaining 19�6% was equal to 165. The
values were >90% and, therefore, the test was very accu-
rate. During training, the neural network correctly pre-
dicted 83% of the cases and 81�8% during the test.
To determine the optimal series of variables to predict
the expected result among the ones used in the input, a net-
work sensitivity analysis was conducted. This analysis pro-
vided the relative importance of the variables included to
determine the output, where 0 indicated a variable with no
predictive validity and 1 indicated a variable with the maxi-
mum predictive validity.
If we consider the importance of the independent vari-
ables of the four dimensions (clarity, attention, terminology,
listening), we can observe the variables that played the most
important role in the four dimensions (Figure 5). The age
variable played an important role in all of the four dimen-
sions: the ability to listen, pay attention, use precise termi-
nology and clarity increased with age. Then followed
seniority, which is also related to age. Gender and level of
education did not have a statistically significant impact on
the effectiveness of communication. This was because
the data on the level of education did not allow its role in
communication to be checked across the years in each
professional. In fact, only education focusing on the com-
munication between health professionals and patients could
have had a strong impact on the SBAR scores.
Discussion
To our knowledge, ANNs have never been used to predict
and prevent harmful nurse–patient communication interac-
tions to increase patient safety. We found that ANNs were
able to correctly predict more than 80% of the communica-
tion characteristics including the communication risks.
Answering to our two research questions, through ANNs
we found that the nurses’ personal characteristics that
enable to predict the risk of communication failures were
age and years of working seniority and experience in the
Emergency Department. Gender and education did not have
a significant impact on communication. In the case of edu-
cation, this would have had a significant impact on commu-
nication between nurses and patients only if education
focused specifically on communication. With regard to the
four communication dimensions, clarity and attention
weighed most in predicting the risk of communication fail-
ures.
In summary, all this means that a correct prediction of
communication failures needs to be followed by appropriate
educational interventions. Notwithstanding, the positive
results that emerged from this study, there is still a long
way to go in this direction because of the many limits.
Limitations
Some of the variables we used were not appropriate to be
specifically processed with the neural networks and there-
fore did not produce any information, so we had to collect
other data, to limit subjective evaluations during observa-
tions. More specifically, with regard to the observers, we
identified the absence of a variable that enabled to under-
stand whether the observer had accessed the personal file of
the person observed, and the opportunity to create the
observer’s profile (personal data, education, scale rating
objectiveness/trustworthiness/level of discrimination or
something similar) to cross-check the data collected by the
observers.
With regard to the population observed, the personal
data needed to be more detailed. For instance, the ‘age of
the nurse observed’ was reported by the observer using a
very wide age bracket. For instance, data on the nurses’
marital status were missing, and the years of education and
incentives for continued education in the field of communi-
cation, including university courses.
Structure30%
25%
20%
15%
10%
5%
0%
Prof. Qualif.
Education
Seniority
Clarity Attention Terminology Listening
Gender
Age
Figure 5 The predictive performance of the independent variable.
Source: Own elaboration.
© 2015 John Wiley & Sons Ltd 9
JAN: ORIGINAL RESEARCH: EMPIRICAL RESEARCH – QUANTITATIVE Risk prediction in ED using ANNs
The variable ‘years of service’ did not include how many
years of work were spent on a given facility or in general
as a nurse. It would be better to state the number of years
of nursing practice, the number of years worked in a given
clinical facility and in the emergency department where the
survey was conducted.
The variable ‘context’ could not be included, because it
was not stated in all of the observations. We also found
that it was important to differentiate the observations made
on admission or triage from those made on discharge.
Another variable that was missing was the one indicating
the nurses’ level of burnout, because it could be important
to reduce the level of subjectivity of the survey on the effi-
cacy of communication in the Emergency Department. We
also found that it would be important to follow the same
patient from admission to discharge, to learn more about
the opinions of the patients and of those take care of them,
because this would allow to control the variables related to
aspects such as their emotional status and concerns.
With regard to ANNS, for a specified data set there may
be an infinite number of relevant network structures that
can learn the characteristics of the data. The question
remains: what size of network will be best for a specific
data set. Unfortunately, it is not easy to answer this ques-
tion. The neural network structure that offers the best
result for a specific problem needs to be experimentally
decided. The size, the characteristics of the training data set
and the number of iterations are the other factors that
affect the generalization (and therefore predictive) capabili-
ties of a neural network. Generally, the network size influ-
ences its complexity and ‘learning time’ but, most
importantly, it affects the generalization capabilities of the
network. Generalization is the ability of the neural network
to interpolate and extrapolate data that it has not seen
before. The predictive power of an ANN depends on how
well it generalizes from the training data.
Conclusion
Nurses can play a very important role in healthcare organi-
zations as safety nets, but very often the environment where
they work harbours many latent risks, such as communica-
tion failures. The proactive analysis of the factors that con-
tribute to communication failures not only between health
professionals and patients but also among health profes-
sionals, could be described by the type of organizational
culture and in addition stimulates reflection about the
health professionals’ communication skills and about the
factors linked to the working environments that have an
influence on communication errors.
This study has shown the potential of the artificial neural
network in predicting the risk of communication failures in
Emergency Departments. The model was developed on
some selected input variables from the SBAR data. It
achieved an accuracy of over 80%, which shows the poten-
tial efficacy of an Artificial Neural Network as a prediction
tool. A ‘spin-off’ result of this study was the validation of
Guilbert’s observation grid as a tool to predict communica-
tion failures. Early detection of communication failures in
health professionals who have specific social and cultural
characteristics would allow health organizations to arrange
specific courses only or prevalently for those who have
weak communication skills.
Funding
This research received no specific grant from any funding
agency in the public, commercial or not-for-profit sectors.
Conflict of interest
No conflict of interest was declared by the authors in rela-
tion to the study itself. Note that Editor-in-Chief Roger
Watson is a visiting professor at the University of Genoa
and has been consulted about this paper but, in line with
usual practice, this article was subjected to double-blind
peer review and was edited by another editor.
Author contributions
All authors have agreed on the final version and meet at
least one of the following criteria [recommended by the IC-
MJE (http://www.icmje.org/recommendations/)]:
• substantial contributions to conception and design,
acquisition of data or analysis and interpretation of
data;
• drafting the article or revising it critically for important
intellectual content.
References
Ak M., Cinar O., Sutcigil L., Congologlu E.D., Haciomeroglu B.,
Canbaz H., Yaprak H., Jay L. & Ozmenler K.N. (2011)
Communication skills training for emergency nurses.
International Journal of Medical Sciences 8(5), 397–401.
Amato F., L�opez A., Pe~na-M�endez E.M., Va�nhara P., Hampl A. &
Havel J. (2013) Artificial neural networks in medical diagnosis.
Journal of Applied Biomedicine 11, 47–58. doi:10.2478/v10136-
012-0031-x.
Bagnasco A., Tubino B., Piccotti E., Rosa F., Aleo G., Di Pietro
P. & Sasso L. (2013) Identifying and correcting
10 © 2015 John Wiley & Sons Ltd
A. Bagnasco et al.
communication failures among health professionals working in
the Emergency Department. International Emergency Nursing
21, 168–172.
Bagnasco A., Pagnucci N., Tolotti A., Rosa F., Torre G. & Sasso
L. (2014) The role of simulation in developing communication
and gestural skills in medical students. BMC Medical Education
14, 106. doi:10.1186/1472-6920-14-106.
Bardan R.T. (2014) Artificial neural networks. Applications in
urology. Timisoara Medical Journal 54, 289–296.
Barbosa D., Roupar D., Ramos J., Tavares A. & Lima C. (2012)
Automatic small bowel tumor diagnosis by using multi-scale
wavelet-based analysis in wireless capsule endoscopy images.
Biomedical Engineering Online 11, 3. doi:10.1186/1475-925X-
11-3.
Catalogna M., Cohen E., Fishman S., Halpern Z., Nevo U. & Ben-
Jacob E. (2012) Artificial neural networks based controller for
glucose monitoring during clamp test. PLoS ONE 7(8), e44587.
doi:10.1371/journal.pone.0044587.
Champion R., Kinsman L.D., Lee G.A., Masman K.A., May E.A.,
Taylor M.D., Thomas P.R. & Williams R.J. (2007) Forecasting
emergency department presentations. Austrian Health Review
31, 83–90.
Day S., Dalto J., Fox J. & Turpin M. (2006) Failure mode and
effects analysis as a performance improvement tool in trauma.
Journal of Trauma Nursing 13(3), 111–117.
Dey P., Lamba A., Kumari S. & Marwaha N. (2011) Application
of an artificial neural network in the prognosis of chronic
myeloid leukemia. Analytical and Quantitative Cytology and
Histology 33(6), 335–339.
Fernandez de Canete J., Gonzalez-Perez S. & Ramos-Diaz J.C.
(2012) Artificial neural networks for closed loop control of in
silico and ad hoc type 1 diabetes. Computer Methods and
Programs in Biomedicine 106, 55–66.
Flin R., Martin L., Goeters K., Hoermann J., Amalberti R., Valot
C. & Nijhuis H. (2003) The development of the NOTECHS
system for evaluating pilots’ CRM skills. Human Factors and
Aerospace Safety 3, 95–117.
Guilbert J.-J. (1990) Guide p�edagogique pour les personnels de
sant�e. World Health Organization, Geneva, Switzerland.
Guilbert J.-J. (2002) The ambiguous and bewitching power of
knowledge, skills and attitudes leads to confusing statements of
learning objectives. Education for Health 15(3), 362–369.
Haig K.M., Sutton S. & Whittington J. (2006) SBAR: a shared
mental model for improving communication between clinicians.
The Joint Commission Journal on Quality and Patient Safety 32,
167–175.
HebbD.O. (1949)TheOrganization of Behavior. Wiley, NewYork.
Hollander J.E., Blomkalns A.L., Brogan G.X., Diercks D.B., Field
J.M., Garvey J.L., Gibler W.B., Henry T.D., Hoekstra J.W.,
Holroyd B.R., Hong Y., Kirk J.D., O’Neil B.J., Jackson R.E.,
Aufderheide T., Blomkalns A.L., Brogan G.X., Christenson J.,
Collins S., Diercks D.B., Fesmire F.M., Garvey J.L., Green G.B.,
Lindsell C.J., Peacock W.F., Pollack C.V., Zalenski R.;
Multidisciplinary Standardized Reporting Criteria Task Force;
Standardized Reporting Criteria Working Group of Emergency
Medicine Cardiac Research and Education Group-International
(2004) Standardized reporting guidelines for studies evaluating
risk stratification of emergency department patients with
potential acute coronary syndromes. Annals of Emergency
Medicine 44(6), 589–598.
Hu X., Cammann H., Meyer H.A., Miller K., Jung K. & Stephan
C. (2013) Artificial neural networks and prostate cancer-tools for
diagnosis and management. Nature Reviews Urology 10(3), 174–
182. doi:10.1038/nrurol.2013.9.
Irfan Khan Y., Zope P.H. & Suralkar S.R. (2013) Importance of
Artificial Neural Network in Medical Diagnosis disease like
acute nephritis disease and heart disease. International Journal of
Engineering, Science and Innovative Technology 2(2), 210–217.
Jones F., Podila P. & Powers C. (2013) Creating a culture of safety
in the emergency department: the value of teamworking training.
Journal of Nursing Administration 43(4), 194–200.
Jones S.S., Thomas A., Evans R.S., Welch S.J., Haug P.J. & Snow
G.L. (2008) Forecasting daily patient volumes in the emergency
department. Academic Emergency Medicine 15, 159–170.
Kilner E. & Sheppard L.A. (2010) The role of team work and
communication in the emergency department: a systematic
review. International Emergency Nursing 18(3), 127–137.
Lago P., Bizzarri G., Scalzotto F., Parpaiola A., Amigoni A., Putoto
G. & Perilongo G. (2012) Use of FMEA analysis to reduce risk
of errors in prescribing and administering drugs in paediatric
wards: a quality improvement report. BMJ Open 2(6), pii:
e001249. doi: 10.1136/bmjopen-2012-001249.
Landsberger H.A. (1958) Hawthorne Revisited. Cornell University
Press, Ithaca.
Li Y., Rauth A.M. & Wu X.Y. (2005) Prediction of kinetics of
doxorubicin release from sulfopropyl dextran ion-exchange
microspheres using artificial neural networks. European Journal
of Pharmaceutical Sciences 24, 401–410.
Liu B. & Jiang Y. (2013) A multi-target training method for
artificial neural network with application to computer-aided
diagnosis. Medical Physics 40, 011908. doi:10.1118/1.4772021.
McCulloch W.S. & Pitts W.H. (1943) A logical calculus of the
ideas imminent in nervous activity. The Bulletin of Mathematical
Biophysics 5, 115–133.
Meginniss A., Damian F. & Falvo F. (2012) Time out for patient
safety. Journal of Emergency Nursing 38(1), 51–53.
Pandini A., Fraccalvieri D. & Bonati L. (2013) Artificial neural
networks for efficient clustering of conformational ensembles and
their potential for medicinal chemistry. Current Topics in
Medical Chemistry 13(5), 642–651.
Patel A., Mehta T., Patel M., Patel K. & Patel N. (2012) Design
porosity osmotic tablet for delivering low and pH-dependent
soluble drug using an Artificial Neural Network. Current Drug
Delivery 9, 459–467.
Qeethara Kadhim A-S. (2011) Artificial neural network in medical
diagnosis. International Journal of Computer Science 8(2), 150–
153.
Rodriguez Galdon B., Penna-Mendez E., Havel J., Rodriguez E. &
Diaz Romero C. (2010) Cluster analysis and artificial neural
networks multivariate classification of onion varieties. Journal of
Agricultural Food Chemistry 58, 11435–11440.
Rosenblatt F. (1958) The perceptron: a probabilistic model for
information storage and organization in the brain. Psychological
Review 65, 386–408.
Saghiri M., Asgar K., Boukani K., Lotfi M., Aghili H., Delvarani
A., Karamifar K., Saghiri A., Mehrvarzfar P. & Garcia-Godoy F.
© 2015 John Wiley & Sons Ltd 11
JAN: ORIGINAL RESEARCH: EMPIRICAL RESEARCH – QUANTITATIVE Risk prediction in ED using ANNs
(2012) A new approach for locating the minor apical foramen
using an artificial neural network. International Endodontic
Journal 45, 257–265.
Siri A. (2014) Predicting Students’ Academic Dropout Using
Artificial Neural Network. Nova Science Publishers, Inc. U.S.,
Commack, NY.
Spencer R., Logan P. & Coiera E. (2002) Supporting
Communication in the Emergency Department. Centre for
Health Informatics. University of New South Wales, Sidney,
NSW.
Schweigler L.M., Desmond J.S., McCarthy M.L., Bukowski K.J.,
Ionides E.L. & Younger J.G. (2009) Forecasting models of
emergency department crowding. Academic Emergency Medicine
16, 301–308.
Werbos P.J. (1974) Beyond Regression: New Tools for Prediction
and Analysis in the Behavioral Sciences. PhD thesis, Harvard
University, Published in the The roots of backpropagation: from
ordered derivatives to neural networks and political forecasting
(Vol. 1) Werbos, P. J. (1994) John Wiley & Sons. New York.
Wesolowski M. & Suchacz B. (2012) Artificial neural networks:
theoretical background and pharmaceutical applications: a
review. Journal of AOAC International 95, 652–668.
Woods K. & Bower K.W. (1997) Generating ROC curves for
artificial neural networks. IEEE Transactions on Medical
Imaging 16, 329–337.
Xu M., Wong T.C. & Chin K.S. (2013) Modeling daily patient
arrivals at Emergency Department and quantifying the relative
importance of contributing variables using artificial neural
network. Decision Support Systems 54, 1488–1498.
Zhang Q.-J. & Gupta K.C. (2000) Neural Networks for RF and
Microwave Design, Chapter 3. Artech House Inc., Boston, MA,
pp. 61–103.
The Journal of Advanced Nursing (JAN) is an international, peer-reviewed, scientific journal. JAN contributes to the advancement of
evidence-based nursing, midwifery and health care by disseminating high quality research and scholarship of contemporary relevance
and with potential to advance knowledge for practice, education, management or policy. JAN publishes research reviews, original
research reports and methodological and theoretical papers.
For further information, please visit JAN on the Wiley Online Library website: www.wileyonlinelibrary.com/journal/jan
Reasons to publish your work in JAN:
• High-impact forum: the world’s most cited nursing journal, with an Impact Factor of 1·527 – ranked 14/101 in the 2012 ISI Jour-
nal Citation Reports © (Nursing (Social Science)).
• Most read nursing journal in the world: over 3 million articles downloaded online per year and accessible in over 10,000 libraries
worldwide (including over 3,500 in developing countries with free or low cost access).
• Fast and easy online submission: online submission at http://mc.manuscriptcentral.com/jan.
• Positive publishing experience: rapid double-blind peer review with constructive feedback.
• Rapid online publication in five weeks: average time from final manuscript arriving in production to online publication.
• Online Open: the option to pay to make your article freely and openly accessible to non-subscribers upon publication on Wiley
Online Library, as well as the option to deposit the article in your own or your funding agency’s preferred archive (e.g. PubMed).
12 © 2015 John Wiley & Sons Ltd
A. Bagnasco et al.