J Med Syst (2007) 31:547550

DOI 10.1007/s10916-007-9097-5

ORIGINAL PAPER

Backpropagation ANN-Based Prediction of Exertional

Heat Illness
Yogender Aggarwal & Bhuwan Mohan Karan &
Barda Nand Das & Tarana Aggarwal &
Rakesh Kumar Sinha

Received: 14 July 2007 / Accepted: 24 August 2007 / Published online: 14 September 2007
# Springer Science + Business Media, LLC 2007

Abstract Exertional heat illness is primarily a multi-system

disorder results from the combined effect of exertional and
thermoregulation stress. The severity of exertional heat
illness can be classified as mild, intermediate and severe
from non-specific symptoms like thirst, myalgia, poor
concentration, hysteria, vomiting, weakness, cramps, impaired judgement, headache, diarrhea, fatigue, hyperventilation, anxiety, and nausea to more severe symptoms like
exertional dehydration, heat cramps, heat exhaustion, heat
injury, heatstroke, rhabdomyolysis, and acute renal failure.
At its early stage, it is quite difficult to find out the severity of
disease with manual screening because of overlapping of
symptoms. Therefore, one need to classify automatically the
disease based on symptoms. The 7:10:1 backpropagation
artificial neural network model has been used to predict the
clinical outcome from the symptoms that are routinely
available to clinicians. The model has found to be effective
in differentiating the different stages of exertional heat-illness
with an overall performance of 100%.

Y. Aggarwal (*) : B. N. Das : R. K. Sinha

Department of Biomedical Instrumentation,
Birla Institute of Technology,
Mesra, Ranchi, Jharkhand 835215, India
e-mail: yogender.aggarwal@gmail.com
B. M. Karan
Department of Electrical and Electronics Engineering,
Birla Institute of Technology,
Mesra, Ranchi, Jharkhand 835215, India
T. Aggarwal
Department of Medicine, Dr. R. M. L. Hospital,
New Delhi, India

Keywords ANN . Backpropagation . Classification .

Exertional heat illness

Introduction
Exertional heat illness (EHI) encompasses a spectrum of
disorders deriving from the combined stresses of exertional
and thermoregulation [1]. These include exertional dehydration, heat cramps, heat exhaustion, heat injury, heat
stroke, rhabdomyolysis, and acute renal failure. It is quite
difficult or impossible to distinguish these entities during
early phase of EHI and in fact, they overlap and are
differentiated as the clinical manifestations evolve. They
represent primarily a continuum of multi-system illness
related to elevation in body temperature and the metabolic
and circulatory processes that are brought about by exercise
and the bodys thermoregulation response [2].
With the increased volume of information available to
physicians from new medical technologies, the process of
classifying different sets of symptoms under a single name
becomes difficult. Furthermore, a single symptom may be
indicative of several different diseases, as in the case of
EHI. Thus, the automated classification of EHI using
artificial neural network (ANN) is important because
manual screening is a tedious and error frame task. ANN
models are based on a set of multilayered interconnected
equations, which use non-linear statistical analysis to reveal
previously unrecognized relations between given input
variables and an output variable. Several studies have
shown that ANN is accurate and reliable in diagnosis and
outcome prediction in diverse clinical settings [39].
The present work aimed to develop an ANN based
model to predict clinical outcome, by use of symptoms

548

J Med Syst (2007) 31:547550

routinely available to clinicians, in a group of patients

presenting with EHI.

Material and methods

Study design
The present paper has used three layered feed-forward
backpropagation ANN to classify different stages of EHI
(mild, intermediate, and severe) based on clinical symptoms
(Table 1) [2]. The digital values of clinical EHI symptoms
between 01 have been used to build the predictive model.
The network was implemented via software by using C++
programming language on a computer [10]. The network
was trained for ten million iterations with the error
tolerance level and learning rate parameter were assigned
0.001 and 0.01 respectively.
The backpropagation neural network has advantage of
available effective training and better-understood system
behavior [4]. Neurons or nodes in each layer are
interconnected in a feedforward fashion. The connections
between different layers of nodes have associated weights,
which act upon outputs of the first layer of nodes before
they are processed next. The nodes of hidden and output
layers have a transfer function. The sigmoid function was
used in this study.

The weights were adjusted as follows [4]:


Wij t 1 Wij t hd j xi a Wij t  Wij t  1
where Wij(t) is the connection weight from a node i in one
layer at time t, xi is either an input or the output of the
hidden node i, j is an error term for node j, is a learning
rate factor and is a momentum factor (0<<1). If j is an
output node, then





d j y j 1  y j dj  y j
where dj is the desired output of node j and yj is the actual
output. If node j is hidden node, however, the computation
of error terms becomes


X
d j xj 1  xj
d k Wjk
k

where k is summed over all nodes in the layer above node j.

Hidden node thresholds were adjusted in a similar way by
assuming they are connection weights on links from
auxiliary constant valued inputs.
Estimation of the EHI
In the three-layered ANN configuration, digital values of
the clinical EHI symptoms were used as the input vector to
the ANN. With various experimentations, optimized neural
network architecture 7:10:1 was obtained both for training

Table 1 Clinical variables used to build the predictive model

Serial
no.

Symptoms

Mild EHI
Pathological condition

Intermediate EHI
Pathological condition

Severe EHI
Pathological condition

Non-specific

Orthostatic

Exertional syncope

Sustained hypotension and

cardiovascular collapse

Encephalopathy

Thirst, myalgia, poor concentration,

hysteria, vomiting, weakness, cramps,
impaired judgement, headache, diarrhea,
fatigue, hyperventilation, anxiety, nausea
Faintness, dizziness, wobble legs, stumbling
gait, blurred vision, tunnel vision,
scotomate, blackout, collapse (without loss
of consciousness
Lethargy, drowsiness

Hyperthermia

100F<max. temp.<104F

Exertional heat stroke/coma, CNS

signs, seizure, speech ataxia
106F<max. temp.

Muscle and
rhabdomyolysis

Exertional heat cramps (muscular pain)

6
7

Renal
Dehydration

Dry oral mucous membrane, dry tongue,

tenting

Exertional heat injury and

amnesia
104F<max. temp.
<106F
Mild rhabdomyolysis/
muscle weakness, pain,
tenderness, stiffness,
myoglobinuria
Renal insufficiency
Hypotension, syncope,
lethargy, fatigue, poor
concentration

Severe rhabdomyolysis/muscle
numbness, weakness, pain,
tenderness, stiffness, myoglobinuria
(with acute renal failure)
Acute renal failure
Confusion, obtundation, or coma

J Med Syst (2007) 31:547550

549

and testing purpose. During training, a training file

TRAINING.DAT with the digital values of EHI symptoms along with an output was presented. The error
tolerance, number of training iterations and learning rate
parameter were assigned as 0.001, 10 millions and 0.01,
respectively. Once the simulator reaches the error tolerance
specified or achieved the maximum number of iterations,
assigned for training, the simulator save the state of
network by saving all its weights in WEIGHT.DAT file.
The output values have been compared to the desired
outputs and if the outputs have found same as desired
outputs, no further training is required. Otherwise weights
have to adapt further by training with least mean square
rule. Obtaining the error in minimum possible value, results
in absolute correction. During testing, the testing file
TEST.DAT with digital values of clinical EHI symptoms
are presented to the ANN, which read the input and finds
the output of the network.
Testing protocol
Our sample size consist of 60 data sets (9 training sets and
51 test sets) described by seven normalized diagnostic
attributes belonging to three different classes (mild,
intermediate, and severe) and a single neuron in output
layer with ten neurons in the hidden layer.

Results and discussion

Before applying the ANN for the classification of EHI, the
network needs to be optimized to get optimum performance. Different learning rates were investigated in the
range of 0.010.5. Table 2 shows the effects of different
learning rates on the performance of the ANN with the
structure of 7101 (nodes of inputhiddenoutput layers).
The best performance was when the learning rate was
chosen as 0.01 with which an overall accuracy of 98.03%
was obtained for testing sets. The effects of number of
hidden nodes are presented in Table 3. With a fixed number

Table 2 Effects of learning rate on effectiveness of 7:10:1 network

(no. of iterations=106)
Learning rate

Accuracy % test pattern

0.01
0.05
0.1
0.2
0.5

98.03
78.43
82.35
84.31
78.43

Table 3 Effects of number of hidden layer nodes on performance of

ANN (Testing results)
Hidden nodes

Accuracy % test pattern

2
5
10
13
15
20

96.07
94.11
100
98.03
98.03
98.03

of iterations (106), ten hidden nodes resulted in the best

performance compared with other combinations of hidden
nodes. With the 7101 structure and optimized learning
rate value (0.01), the ANN identified the 51 EHI test sets
with an accuracy of 100%.
In EHI classification, there are no clinically useful
predictive techniques available to assist the physician in
the diagnosis of patients. In this study, we used ANN based
modeling techniques is an attempt to predict the severity of
EHI (mild, intermediate, and severe). For any predictive
technique to be useful in making a diagnostic decision, an
important feature is that only data that are readily available
to the clinician at the time of diagnosis are used [9]. Our
model used only data derived from clinical history and
physical examination. We emphasize that our model is not
meant to replace or substitute for an experienced clinician;
on the contrary, we suggests that the model should be
viewed as a decision aid for the busy emergency department particularly in times heat waves in summer. The result
of this study are promising, suggesting that a system based
on this approach will be accurate and user friendly, while
simple enough to be implemented at low cost. The present
report shows overall 100% sensitivity in classification of
EHI. Input layer of ANN was designed with seven most
identified clinical data. However, the addition of any new
clinical symptom and further training and testing of ANN
can alter the recognition pattern. In the current investigation
all 60 data sets for heat illness cases were recorded and
provided by expert clinicians. The outcome of the ANN
was also tested clinically for these data sets and it can be
presumed that this architecture with 7:10:1 can provide a
platform to develop an automated system to aid clinicians
and researchers.
It has already been established that the success of ANN
in classification involves the optimization of the network
structure and the parameters. In the present study, various
combinations of three-layered backpropagation network
were tested with assigning different learning rate parameters and the most reliable performance rate was derived

550

with the ANN configuration of 7:10:1. However, potential

clinical studies are now needed to assess the ability of
ANN model to improve the patient management and
healthcare resources in clinical practice and in diverse
population.

References
1. Varghese, G. M., John, G., Thomas, K., Abraham, O. C., and
Mathai, D., Predictors of multi-organ dysfunction in heat stroke.
Emerg. Med. J. 22:185187, 2005.
2. Gardner, J. W., and Kark, J. A., Clinical diagnosis, management,
and surveillance of exertional heat illness. In: Pandolf, K. B., and
Burr, R. E., (Eds), Medical Aspects of Harsh Environments.
Chapter 7, Vol. 1, pp. 231279, 2002.
3. Sinha, R. K., Aggarwal, Y., and Das, B. N., Backpropagation
artificial neural network detects changes in electro-encephalogram power spectra of syncopic patients. J. Med. Syst. 31:6368,
2007.

J Med Syst (2007) 31:547550

4. Sinha, R. K., Aggarwal, Y., and Das, B. N., Backpropagation
artificial neural network classifier to detect changes in heart sound
due to mitral valve regurgitation. J. Med. Syst. 31(3):205209,
2007.
5. Sinha, R. K., An approach to estimate EEG power spectrum as an
index of heat stress using backpropagation artificial neural
network. Med. Eng. Phys. 29:120124, 2007.
6. Sinha, R. K., Electro-encephalogram disturbances in different
sleepwake states following exposure to high environmental heat.
Med. Biol. Eng. Comput. 42:282287, 2004.
7. Sinha, R. K., Ray, A. K., and Agrawal, N. K., An artificial neural
network to detect EEG seizures. Neurol. India 53:399400, 2004.
8. Sinha, R. K., Artificial neural network detects changes in electroencephalogram power spectrum of different sleep-wake states in
an animal model of heat stress. Med. Biol. Eng. Comput. 41:595
600, 2003.
9. Das, A., Ben-Menachem, T., Cooper, G. S., Chak, A., Jr. Sivak,
M. V., Gonet, J. A., and Wong, R. C. K., Prediction of outcome in
acute lower-gastrointestinal haemorrhage based on an artificial
neural network: Internal and external validation of a prediction
model. The Lancet 362:12611266, 2003.
10. Rao, V., and Rao, H., C++ neural networks and fuzzy Logic, 2nd
edn., New Delhi, India: BPB Publications, pp. 123176, 1996.