Artificial Neural Network and Its Application in Medical Disease

Prediction: Review Article

Putri Alief Siswanto

and Riries Rulaningtyas

Biomedical Engineering Graduate Program, Faculty of Science and Technology, Airlangga University,

Surabaya, Indonesia

Department of Physics, Faculty of Science and Technology, Airlangga University, Surabaya, Indonesia

Keywords: Artificial Neural Network, Prediction, Regression.

Abstract: Artificial Neural Network (ANN) has gained considerable attention in many fields, including medicine.

Despite this, many physicians are still oblivious to the concept of ANN. In this paper, therefore, we attempt

to provide an overview of the features of ANN. In addition, performance comparison in medical disease

prediction between ANN and regression as a standard statistical technique is also conducted. This is due to

the proficiency of ANN in modeling complex, non-linear relationships which makes it highly attractive for

diagnostic or prognostic purposes. It can be concluded that ANN has similar performance to, and in some

cases, is better than regression.

1 BACKGROUND

In the last few decades, artificial intelligence (AI) has

gained increasing popularity among researchers. AI is

a new technology in computer science and is

distinguished from conventional computers. Whereas

its counterpart could only perform specific tasks

based on instructions given or often referred as

program (software), AI tries to adopt the flexibility of

the human brain (Sazli, 2006).

One of the most popular branches of AI is that of

artificial neural networks (ANN). ANN is an

information processing system that resembles

biological neural networks (Haykin, 2009); ANN has

been used extensively in many disciplines, including

medicine (Ramesh et al., 2004). Despite its potential,

ANN is nevertheless still unknown to many

practitioners. Therefore, we will provide an overview

of ANN that includes definition, components,

architecture, training methods and the

backpropagation method. In the future, it is expected

that not only will more physicians become familiar

with this method, but that they will also be able to

properly apply this method.

The application of ANN is mostly used for

prediction and modeling. In order to provide a more

comprehensive understanding of ANN’s application

in medical disease prediction specifically,

comparative studies of ANN and standard statistical

analysis are also discussed.

2 METHOD

The literature search was conducted using the terms

‘neural network’, ‘prediction’ and ‘regression’ from

the year 2000 to the present. Comparison studies were

limited to medical applications. Titles and abstracts

were reviewed to identify papers that could contribute

to this discussion, and full-text versions were

acquired.

3 DISCUSSION

3.1 Definition of ANN

ANN is a mathematical representation of the human

neural architecture, and reflect its learning and

generalization abilities (Dave and Dutta, 2014). In

ANN, information processing occurs in elements

called neurons. Signals are transmitted from one

neuron to another through connection. Each

connection has a synaptic weight that is multiplied

with the input signal. Each neuron then applies an

Siswanto, P. and Rulaningtyas, R.

Artiﬁcial Neural Network and Its Application in Medical Disease Prediction: Review Article.

DOI: 10.5220/0009387400170025

In Proceedings of the 4th Annual International Conference and Exhibition on Indonesian Medical Education and Research Institute (The 4th ICE on IMERI 2019), pages 17-25

ISBN: 978-989-758-433-6

activation function to the input signal to determine

output (Nicoletti, 2000).

ANN emulates the way the human brain learns

through examples. The brain makes some

adjustments to synaptic connections during the

learning process. This is applied to ANN. During the

learning phase, weights are adjusted to values such

that each input produces desired output (Huang,

2017).

3.2 ANN Architecture

Neurons are organized in layers. Each neuron in a

layer is connected with every other neuron in the

previous or next layer through a synaptic weight. The

values of this weight indicates the strength of the

connection (Richards, 2006).

The structure of the neural network is formed by

an input layer, one or more hidden layers and the

output layer. Network architecture plays a major role

in the success of ANN. Therefore, an optimal network

must be determined (Amato et al., 2012).

The input layer acts as an intermediary between

the network and the outside. Neurons in this layer

only forward data to the following layer, as there are

no data changes being made (Wang et al., 2009).

Information received by neurons in the input layer

are then transfered to the output layer through the

hidden layer; the hidden layer helps network to

recognize still more patterns. The number of hidden

layers depends on the complexity of the system

studied (Sazli, 2006). For most of the biomedical

applications, there is no evidence that suggests that

more than one hidden layer will meaningfully add to

the predictive capabilities of a network (Ahmed,

2005).

The selection of a number of hidden neurons

remains one of the major problems in ANN, since it

is dependent on factors unique to each case. Too

many hidden neurons will cause the network to

overfit, or lose its generalizing ability, since the

network tends to remember all the given examples.

Conversely, if the number of hidden neurons is

smaller than the complexity of the problem, it will

cause the network to underfit and hinder convergence

(Ramli and Clean, 2011)

Convergence is a condition that happens when the

network shows minimal errors after the training

process.The network will stop learning when it

reaches convergence (Sun et al., 2003). Convergency

of target and model output values can be observed

with the use of mean error (BIAS), mean absolute

error (MAE), mean absolute percentage error

(MAPE) and root mean square error (RMSE) (Bilgili

and Sahin, 2010).

Panchal and Panchal (2014) conducted a literature

review regarding the method of selecting a number of

hidden neurons. The hidden layer has significant

influence on the output, although it does not interact

outside of the network. Several approaches are

available to determine the number of hidden neurons,

such as trial and error, rule of thumb and the simple

method. It should be noted that researchers who

proposed some of the methods did not use the results

of the formula in the study they conducted. Therefore,

it is recommended to use trial-and-error methods as

long as there is no accurate guideline regarding the

selection of the number of hidden neurons.

If the network has several hidden layers, the

lowest layer functions to receive input. The value of

the input (Nett) of the j-th neuron in the hidden layer

equals to the number of multiplications between the

weight value (𝑊) and the output value (𝑂) of the i-th

neuron in the previous layer (input neuron) plus bias

(W, jth neuron).

𝑁𝑒𝑡𝑡



𝑊



𝑖𝑗



.𝑂



𝑖



𝑊𝑜𝑗

(1)

Bias is the value of weight which connects bias

neurons and neurons in the next layer. The input of

bias neuron always has a value of 1. Bias neurons can

be added to input layer and hidden layer. Bias can be

initialized randomly within a range of -1 to 1. Bias is

aimed to improve the network’s output to

approximate the desired value (Abiodun et al., 2018).

The output value of neurons in the hidden layer is

a function of the input value f (net(j)). Several choices

for activation functions are available. The same

formulas used in the hidden layer are applied to the

output layer, including the activation function, but

output are considered to be the result of the learning

process (Liang et al., 2010).

There are three types of ANN architecture: single-

layer, multi-layer and competitive-layer. A single-

layer network has a set of neuron inputs directly

connected to a set of outputs, and no hidden layers are

found (Alsmadi et al., 2009). This results in limited

ability while generalizing more complex problems.

To overcome this, a single-layer network adds one or

more hidden layers, or a so-called multi-layer. A

competitive-layer network is similar to a single- or

multi-layer network, but there is feedback from

output neurons towards input throughout the network

(feedback loop) (Nicoletti, 2000).

The 4th ICE on IMERI 2019 - The annual International Conference and Exhibition on Indonesian Medical Education and Research Institute

Figure 1: Example of single-layer ANN.

Figure 2: Example of multi-layer ANN.

Figure 3: Example of competitive-layer ANN.

3.3 ANN Components

The basic building block of neural networks is the

neuron. The neuron can be perceived as a processing

unit. In a neural network, neurons are connected with

each other through synaptic weights. Weight stores

information at certain values. Positive values will

strengthen the signal and vice versa.

Each neuron receives weighted information via

these synaptic connections and then passes the

weighted sum of those input signals in order to be

compared with a threshold value using activation

function. If the amount of input exceeds a certain

threshold value, the neuron will be activated and send

its output. If the amount of input is smaller than the

threshold value, the information will not be passed on

(Nicoletti, 2000).

Figure 4: Structure of ANN.

For example, if 𝑥



,𝑥



,…𝑥



is an input neuron and

𝑤



,𝑤



,…𝑤



is synaptic weight to the output

neuron, then the weighted sum will give an output of:

𝑢



𝑥



𝑤



𝑥



𝑤



⋯𝑥



𝑤



(2)

Weights can be randomly initialized in a range value

of -1 to 1 and should be a small value.

The activation function is a function that maps

input to its output. It is not only used to determine the

output of a neuron, but also to modify the output into

a certain range of values so that it can approach

maximum and minimum values, as well (Siang,

2005).

There are several types of activation functions:

the step biner, bipolar, linear and sigmoid functions.

Must-have characteristics for activation functions are

continuous, differentiable, not asymptotic, not

monotonically decreasing and easily calculated

(Richards, 2006).

The step biner function converts input value into

binary output (0 or 1). The bipolar function has an

output result of 1, 0 or -1. A linear function has the

same output value as input value. The sigmoid

function is often used in a backpropagation algorithm

due to its shape of function and easily computed

derivatives. Parameter related to the shape (slope) of

the sigmoid function is alpha (α). The greater the

value, the more the non-linear sigmoid function will

produce.

There are two steps of the ANN: the training and

testing phase. Each step uses different data sets. The

training phase produces the final weight which is later

used in the testing phase (Toshniwal and Alsma,

2009).

There is no exact rule regarding the composition

of training and testing data. Crowther and Cox (2005)

divided training and testing data with a distribution of

70:30 of the total data. Meanwhile, some researchers

divide data with a composition of 80:20 for training

Artiﬁcial Neural Network and Its Application in Medical Disease Prediction: Review Article

and testing data. Randomization is advised before

dividing the data in order to reduce bias.

There are two important aspects of the training

data set. First, network performance not only depends

on network parameters, but also on the accuracy of

the training data set. Training data should be able to

represent all the inputs which are intended to be

trained in order to precisely produce output (Azmi

and Cob, 2010). Extreme values in data can be well

represented if included in training data, but the

network cannot accurately extrapolate values if it is

outside the training data (Ramli and Clean, 2011).

Second, the data needs to be normalized before being

processed. This aims to simplify the network during

the training phase. Network output is then

denormalized to return to its original value (Jiang et

al., 2010).

In addition to training and testing data, there is

validation data. The validation step is used to

optimize model parameters and monitor network

performance (Lisboa, 2002). Performance of training

and validation data sets will improve during the

training process. But if network capacity is greater,

then performance of validation data set will decrease.

This can be a sign of overlearning or overfitting, and

the training process should be immediately stopped

(early stopping) (Ahangar, 2010).

In general, training data is divided into training

and validation data sets. This validation technique is

called the split-sample method (Park and Han, 2018).

Another popular validation method is n-fold cross-

validation. This method is performed by dividing

training data sets into n data sets. The network is

validated every time the learning process is

completed using different subsets. The training

process is carried out with remaining data sets which

were not being used for validation. Model

performance is measured by calculating the average

RMSE of all validations performed (Almeida, 2002).

Other validation techniques are temporal and

external validation (Park and Han, 2018). Temporal

validation is carried out using the cohort at different

times (Lisboa and Taktak, 2006). External validation

involves a larger database (apart from previously

tested population) (Terrin et al., 2003), which is

usually conducted through multicenter studies.

Testing on a separate data set would have provided an

unbiased estimate of the generalization error

(Ahangar, 2010).

Measures of a network’s performance for

diagnostic or predictive models include discrimina-

tion and calibration. Discrimination determines how

well groups within data are separated. Discrimination

performance is typically measured in terms of

sensitivity, specificity and receiver operating

characteristic (ROC) curve analysis. Calibration

refers to correspondence between predicted and

actual probabilities. It is measured by calculating

accuracy or difference between target and output

value (Park and Han, 2018)

3.4 ANN Training Methods

There are three major categories of the training

method. The first is supervised learning, in which a

network is provided with expected output and trained

to respond correctly through error correction and

weight adjustment. This method is commonly used

for pattern/memory association and regression

analysis/function estimation. Some examples of

applications are pattern recognition (face/object

identification), sequence identification (gestures,

language, handwriting) and decision-making (Ashok

et al., 2010). ANN methods belonging to supervised

learning are Hebb’s rule, Perceptron, the Delta rule,

backpropagation, radial basis function (RBF), and

bidirectional associative memory (BAM) (Ramli and

Clean, 2011).

Second is unsupervised learning, in which the

network is provided with no knowledge of expected

output beforehand and trained to discover structures

in presented inputs. Weights are categorized into a

certain range of values in which nearly same input

will produce same output. It is commonly used for

grouping or classification. ANN methods belonging

to unsupervised learning are self-organizing maps

(SOM) and k-means clustering (KNN) (Kusumadewi,

2006).

The third is reinforcement learning, in which a

network is not provided with explicit output; instead,

it is periodically given performance indicators which

act as a supervisory system (Nayak et al., 2001).

3.5 ANN Backpropagation

The most common type of ANN used is the

backpropagation learning algorithm. The basic theory

of backpropagation is based on the error-correction

learning rule, which uses the error function in order

to modify the connection weights to gradually reduce

the error. Error is the difference between actual output

and the desired output. Therefore, backpropagation is

a supervised learning paradigm because it requires a

desired output (Alsmadi et al., 2009).

There are several methods of weight updates,

including gradient descent, quasi newton (LM) and

genetic algorithm. Gradient descent is used in

The 4th ICE on IMERI 2019 - The annual International Conference and Exhibition on Indonesian Medical Education and Research Institute

backpropagation and has better performance

compared to other methods (Ghaffari et al., 2006).

Gradient descent technique is the most common

approach that is utilized in backpropagation. It is

conducted by modifying the weight via values

proportional to the derivative of the error function

(Abiodun et al., 2018).

The disadvantage of the gradient descent method

is that it is likely to get caught up in local minima.

Local minima are when a network achieves an error

that is lower than the surrounding possibilities, but it

is not the smallest possible one. Training that takes

much time but does not give better results is a sign of

local minima.

Backpropagation can be implemented in two

modes, the batch and incremental modes (Aria,

2003). In batch mode, the network weights are

updated once after the presentation of the complete

training set. While in incremental mode, weights are

updated after each pattern presentation. The last mode

is the most frequently used (Sazli, 2006). According

to Ghaffari et al. (2006), there is no significant

difference between the two modes in accuracy, but

batch mode has 3-to-4-times higher convergence

speed than incremental mode.

Backpropagation algorithm can be decomposed in

the following three steps: (a) computation of a feed-

forward stage, (b) backward propagation of the error

to both the output layer and the hidden layers, and (c)

weights update. The algorithm is stopped when the

value of the error function has become less than a

predefined acceptable tolerance (Laudon et al., 2007).

During the feedforward step, each input neuron

) will receive a signal input and forward to each

neuron in hidden layer (Z

). Each neuron in the hidden

layer then calculates its activation and sends a signal

) to each output neuron. Then, each output neuron

) also calculates its activation (y

) to produce a

response to the input given by the network.

Each neuron output then compares its activation

) with a target value (t

) to calculate the error.

Based on that error, 𝛿



is calculated. Factor 𝛿



is used

to distribute errors back to the previous layer.

Meanwhile, factor 𝛿



is used to update weight, which

connects hidden and input layer.

The learning rate is a factor that controls the speed

of the training phase. It is determined through trial

and error to generate the fastest iteration in achieving

convergence (Amin et al., 2012). The input of

learning rate is within range of 0 to 1 (Wilson, 2001).

A too-high value can cause the algorithm to become

unstable due to oscillation of weight update and result

in less generalizing ability. Meanwhile, a too-small

value will result in a longer time to achieve

convergence and lead to local minima (Suliman and

Zhang, 2015).

A maximum epoch is the maximum number of

epochs allowed during the training process. Iteration

will be stopped if epoch in the process exceeds

maximum defined training epochs.

Iteration can also be stopped if network error is

equal to or smaller than the target error. The target

error is a predetermined maximum error value of

network performance. In addition, the training phase

can be stopped if there is no significant weight update

after a long-term process (Ahangar, 2010).

3.6 Application of ANN in Medical

Disease Prediction

In the last decade, the use of AI has become widely

accepted in medical application. This is manifested

by an increasing number of medical devices currently

available on the market that are equipped with AI

features, and also the number of publications in

medical journals (Gant et al., 2001).

ANN is the most well-known method in AI

(Meengoen et al., 2017). Considerable attention has

been paid to the development of ANN and its

application in many areas, including medicine. It is

currently an issue of great interest, especially with

regard to diagnostic or predictive analysis (Amato et

al., 2012). ANN models have been shown to be

valuable tools in reducing the workload on clinicians

by providing decision support (Gant et al., 2001).

Meanwhile, other branches of AI, such as fuzzy

expert systems, evolutionary computation and hybrid

intelligent systems, can be applied to certain clinical

conditions to a limited extent (Ramesh et al., 2004).

ANN models possess such characteristics as:

(a) nonlinearity, which can approximate any non-

linear mathematical function (this is particularly

useful when the relationship between the variables is

too complex and therefore difficult to handle

statistically); (b) noise-insensitivity enables them to

accept a certain amount of uncertain data or

measurement errors without a serious effect on

accuracy; (c) learning and adaptability allows a

network to modify its internal structure (Paliwal and

Kumar, 2009); (d) high parallelism implies fast

processing; and (e) generalization enables application

of the model to untrained data (Ahmed, 2005).

Advantages of using ANN include: (a) less formal

statistical training to develop, (b) having a wide

variety of training algorithms and (c) the ability to

detect complex nonlinear relationships between

variables (Sargent, 2001).

Artiﬁcial Neural Network and Its Application in Medical Disease Prediction: Review Article

Meanwhile, disadvantages of ANN include:

(a) as they are considered to be ‘black box’ methods,

one cannot exactly understand which interactions are

being modeled in their hidden layers (Mantzaris et al.,

2008); (b) weights produced during the training phase

are lacking in interpretability, and therefore

inferences regarding the significance of certain

variables cannot be drawn (Paliwal and Kumar,

2009); (c) model development is empirical (through

trial and error) and requires lengthy development and

time to optimize, so many methodological issues

therefore remain to be solved (Prieto et al., 2016); (d)

models are susceptible to overfitting; and (e) they are

difficult to apply in certain fields because of

computational requirements (Terrin et al., 2003).

Although there is currently ready-to-use software, it

has not yet been standardized (Sargent, 2001).

There are still only a few published clinical trials

for ANN (Lisboa and Taktak, 2006). Further clinical

trials which are properly designed are needed before

this method finds application in a real clinical setting

(Ramesh et al., 2004).

Neural networks can be applied for complex

pattern recognition, classification problems and

function estimation, but are mostly used for

modelling and prediction. The non-linear structure of

ANN enables us to model complex functional forms

(Paliwal and Kumar, 2009). This suggests that ANN

is a promising method for overcoming non-linear

problems demonstrated in biological systems

(Almeida, 2002).

Prediction models in medicine are used for

diagnostic and prognostic purposes. Research has

been conducted on ANN in medicine intended to

improve the accuracy of clinical predictions (Escobar

et al., 2016). The most common method used for

medical disease prediction is the regression model

(Sargent, 2001).

Regression methods have become standard due to

their simple computation (there are many available

computer softwares to fit these models) and

interpretability of model parameters, and in spite of

their limitations. For example, a regression can only

examine linear relationship between variables;

therefore, they may not be able to provide an accurate

prediction in complex non-linear data. Regression

models also need to address many assumptions (such

as normality) before models can be constructed, or

they will produce significant errors. These cause

regression to be inefficient. However, regression

remains the primary choice if one is trying to examine

possible causal relationships among variables

(Eftekhar et al., 2005).

The standardized method is commonly used as a

benchmark while evaluating performance of ANN

models (Lisboa and Taktak, 2006). Different results

were found in studies that performed a comparison of

ANN and regression in medical disease prediction

(Ciampi and Zhang, 2002).

Hassanipour et al. (2019) conducted a systematic

review of 10 studies that predicted trauma-related

outcomes. Results showed that AUC for ANN and

regression were 0.91 (95% CI 0.89–0.83) and 0.89

(95% CI 0.87–0.90), while accuracy for ANN and

regression were 90.5 (95% CI 87.6–94.2) and 83.2

(95% CI 75.1–91.2). These results showed that ANN

had a better performance than regression in predicting

outcomes for trauma patients.

Ture et al. (2005) predicted the risk of

hypertension, while Remzi et al. (2003) predicted the

risk of prostate cancer designed for screening. Both

studies demonstrated that ANN produced more

accurate predictions.

Ottenbacher et al. (2004) comparedthe

performance of ANN and regression in predictive

models for epidemiological research, and no

significant differences were found. The same results

were also obtained in studies performed by Gaudart

et al. (2004), Song et al. (2005) and Behrman et al.

(2007).

According to a meta-analysis by Sargent (2001),

wherein the ANN and regression for medical disease

prediction in 28 studies was compared, ANN had a

similar performance in 50% of the cases, and in 36%

of the cases, ANN had a better performance. In a large

study (sample size > 5000), neither method

dominated the other, but in moderate-sized data (200–

2000), ANN tended to be equivalent to or outperform

regression.

4 CONCLUSION

Artificial Neural Network (ANN) emulates the

structure and function of the human nervous system.

ANN is mostly known for its capability in modeling

complex non-linear problems, such as those

demonstrated in medicine. In comparison to the

regression method, ANN has similar or better

performance. Therefore, ANN is a promising method

to overcome non-linearity commonly found in

medical disease prediction.

The 4th ICE on IMERI 2019 - The annual International Conference and Exhibition on Indonesian Medical Education and Research Institute

REFERENCES

Abiodun, O. I., Jantan, A., Omolara, A. E., Dada, K. V.,

Mohamed, N. A., and Arshad, H. (2018). State-of-the-

art in artificial neural network applications: A survey.

Heliyon, 4(11), e00938. http://doi.org/10.1016/j.heli

yon.2018.e00938.

Ahangar, R. G., Yahyazadehfar, M., and Pournaghshband,

H. (2010). The Comparison of Methods Artificial

Neural Networkwith Linear Regression Using Specific

Variables forPrediction Stock Price in Tehran Stock

Exchange. (IJCSIS) International Journal of Computer

Science and Information Security, 7(2), 38–46.

Ahmed, F. E. (2005). Artificial neural networks for

diagnosis and survival prediction in colon cancer.

Molecular Cancer, 4. http://doi.org/10.1186/1476-

4598-4-29.

Almeida, J. S. (2002). Predictive non-linear modeling of

complex data by artificial neural networks. Current

Opinion in Biotechnology, 13(1), 72–76.

http://doi.org/10.1016/S0958-1669(02)00288-4.

Alsmadi, M., Omar, K., and Noah, S. A. M. (2009). Back

Propagation Algorithm : The Best Algorithm Among

the Multi-layer Perceptron Algorithm. IJCSNS

International Journal of Computer Science and

Network Security, 9(4), 378–383.

Amato, F., López, A., Peña-Méndez, E. M., Vaňhara, P.,

Hampl, A., and Havel, J. (2013). Artificial neural

networks in medical diagnosis. Journal of Applied

Biomedicine, 11(2), 47–58. http://doi.org/10.2478/v10

136-012-0031-x.

Amin, S., Alamsyah, A., and Muslim, M. A. (2012). Sistem

Deteksi Dini Hama Wereng Batang Coklat

Menggunakan Jaringan Syaraf Tiruan Backpropaga-

tion. Unnes Journal of Mathematics, 1(2), 118–123.

http://doi.org/10.15294/ujm.v1i2.1066.

Ashok, V., Rajan Singh, S., and Nirmalkumar, A. (2010).

Determination of Blood Glucose Concentration by

Back Propagation Neural Network. Indian Journal of

Science and Technology, 3(8), 916–98.

http://doi.org/10.17485/ijst/2010/v3i8/29910.

Azmi, M. S. B. M., and Cob, Z. C. (2010). Breast Cancer

Prediction Based on Backpropagation Algorithm. In

2010 IEEE Student Conference on Research and

Development (SCOReD) (pp. 164–168). Putrajaya,

Malaysia: IEEE.

Behrman, M., Linder, R., Assadi, A. H., Stacey, B. R., and

Backonja, M. M. (2007). Classification of patients with

pain based on neuropathic pain symptoms: Comparison

of an artificial neural network against an established

scoring system. European Journal of Pain, 11(4), 370–

376. http://doi.org/10.1016/j.ejpain.2006.03.001.

Bilgili, M., and Sahin, B. (2010). Comparative analysis of

regression and artificial neural network models for

wind speed prediction. Meteorology and Atmospheric

Physics, 109(1–2), 61–72.

Ciampi, A., and Zhang, F. (2002). A new approach to

training back-propagation artificial neural networks:

Empirical evaluation on ten data sets from clinical

studies. Statistics in Medicine,

21(9), 1309–1330.

http://doi.org/10.1002/sim.1107.

Crowther, P. S., and Cox, R. J. (2005). A Method for

Optimal Division of Data Sets for Use in Neural

Networks. In R. Khosla, R. J. Howlett, and L. C. Jain

(Eds.), Knowledge-Based Intelligent Information and

Engineering Systems: 9th International Conference,

KES 2005, Melbourne, Australia, September 14-16,

2005, Proceedings, Part IV (Vol. 3684, pp. 1–7).

Berlin: Springer-Verlag.

Dave, V. S., and Dutta, K. (2014). Neural network based

models for software effort estimation: A review.

Artificial Intelligence Review, 42(2), 295–307.

http://doi.org/10.1007/s10462-012-9339-x.

Eftekhar, B., Mohammad, K., Ardebili, H. E., Ghodsi, M.,

and Ketabchi, E. (2005). Comparison of artificial neural

network and logistic regression models for prediction

of mortality in head trauma based on initial clinical

data. BMC Medical Informatics and Decision Making,

5(1). http://doi.org/10.1186/1472-6947-5-3.

Escobar, G. J., Turk, B. J., Ragins, A., Ha, J., Hoberman,

B., LeVine, S. M., … Kipnis, P. (2016). Piloting

electronic medical record-based early detection of

inpatient deterioration in community hospitals. Journal

of Hospital Medicine, 11(S1), S18–S24. http://doi.org/

10.1002/jhm.2652.

Gant, V., Rodway, S., and Wyatt, J. (2001). Artificial neural

networks: Practical considerations for clinical

application. In R. Dybowski and V. Gant (Eds.),

Clinical Applications of Artificial Neural Networks (1st

ed., pp. 329–356). Cambridge: Cambridge University

Press.

Gaudart, J., Guisiano, B., and Huiart, L. (2004).

Comparison of the performance of multi-layer

perceptron and linear regression for epidemiological

data. Computational Statistics & Data Analysis, 44(4),

547–570. http://doi.org/10.1016/S0167-9473(02)002

57-8.

Ghaffari, A., Abdollahi, H., Khoshayand, M. R., Soltani

Bozchalooi, I., Dadger, A., and Rafiee-Tehrani, M.

(2006). Performance comparison of neural network

training algorithms in modeling of bimodal drug

delivery. International Journal of Pharmaceutics,

327(1–2), 126–138. http://doi.org/10.1016/j.ijpharm.

2006.07.056.

Hassanipour, S., Ghaem, H., Arab-Zozani, M., Seif, M.,

Fararouei, M., Abdzadeh, E., … Paydar, S. (2019).

Comparison of artificial neural network and logistic

regression models for prediction of outcomes in trauma

patients: A systematic review and meta-analysis.

Injury, 50(2), 244–250. http://doi.org/10.1016/j.injury.

2019.01.007.

Haykin, S. O. (2009). Neural Networks and Learning

Machines (3rd ed.). Upper Saddle River, NJ: Pearson

Education, Inc.

Hosseini Aria, E. R., Amini, J. R., and Saradjian, M. R.

(2003). Back Propagation Neural Network for

Classification of IRS-1D Satellite Images. Joint

Workshop of High Resolution Mapping from Space.

Tehran University, Iran.

Artiﬁcial Neural Network and Its Application in Medical Disease Prediction: Review Article

Huang, T.-J. (2017). Imitating the Brain with

Neurocomputer: A “New” Way Towards Artificial

General Intelligence. International Journal of

Automation and Computing, 14(5), 520–531.

http://doi.org/10.1007/s11633-017-1082-y.

Jiang, J., Zhang, J., Yang, G. H., Zhang, D., and Zhang, L.

(2010). Application of back propagation neural network

in the classification of high resolution remote sensing

image: Take remote sensing image of beijing for

instance. In 2010 18th International Conference on

Geoinformatics (pp. 1–6). Beijing, China: IEEE.

Kusumadewi, S. (2006). Jaringan Syaraf Tiruan.

Yogyakarta: Graha Ilmu.

Laudon, K. C., and Laudon, J. P. (2007). Management

Information Systems: Managing the Digital Firm (10th

ed.). Lebanon, IN: Prentice.

Liang, P., Zhaoyang, X., and Jiguang, D. (2010).

Application of BP neural network in remote sensing

image classification. In 2010 International Conference

on Computer Application and System Modeling

(ICCASM 2010) (pp. 212–215). IEEE.

Lisboa, P. J. G. (2002). A review of evidence of health

benefit from artificial neural networks in medical

intervention. Neural Networks: The Official Journal of

the International Neural Network Society, 15(1), 11–

39. http://doi.org/10.1016/s0893-6080(01)00111-3.

Lisboa, P. J. G., and Taktak, A. F. G. (2006). The use of

artificial neural networks in decision support in cancer:

A systematic review. Neural Networks: The Official

Journal of the International Neural Network Society,

19(4), 408–415. http://doi.org/10.1016/j.neunet.2005.

10.007.

Mantzaris, D. H., Anastassopoulos, G. C., and

Lymberopoulos, D. K. (2008). Medical disease

prediction using Artificial Neural Networks. In 2008

8th IEEE International Conference on BioInformatics

and BioEngineering. Athens, Greece: IEEE.

Meengoen, N., Wongkittisuksa, B., and Tanthanuch, S.

(2017). Measurement study of human blood pH based

on optical technique by back propagation artificial

neural network. In 2017 International Electrical

Engineering Congress (iEECON) (pp. 8–10). IEEE.

Nayak, R., Jain, L. C., and Ting, B. K. H. (2001). Artificial

Neural Networks in Biomedical Engineering: A

Review. In S. Valliappan and N. Khalili (Eds.),

Computational Mechanics–New Frontiers for the New

Millennium: Proceedings of the First Asian-Pacific

Congress on Computational Mechanics, Sydney,

N.S.W., Australia, 20-23 November 2001 (Vol. 1, pp.

887–892). Amsterdam: Elsevier.

Nicoletti, G. M. (2000). An Analysis of Neural Networks

as Simulators and Emulators. Cybernetics and Systems,

31(3), 253–282. http://doi.org/10.1080/01969720012

4810.

Ottenbacher, K. J., Linn, R. T., Smith, P. M., Illig, S. B.,

Mancuso, M., and Granger, C. V. (2004). Comparison

of logistic regression and neural network analysis

applied to predicting living setting after hip fracture.

Annals of Epidemiology, 14(8), 551–559.

http://doi.org/10.1016/j.annepidem.2003.10.005.

Paliwal, M., and Kumar, U. A. (2009). Neural networks and

statistical techniques: A review of applications. Expert

Systems with Applications, 36(1), 2–17. http://doi.org/

10.1016/j.eswa.2007.10.005

Panchal, F. S., and Panchal, M. (2014). Review on Methods

of Selecting Number of Hidden Nodes in Artificial

Neural Network. International Journal of Computer

Science and Mobile Computing, 3(11), 455–464.

http://doi.org/10.1155/2013/425740.

Park, S. H., and Han, K. (2018). Methodologic Guide for

Evaluating Clinical Performance and Effect of

Artificial Intelligence Technology for Medical

Diagnosis and Prediction. Radiology, 286(3), 800–809.

http://doi.org/10.1148/radiol.2017171920.

Prieto, A., Prieto, B., Ortigosa, E. M., Ros, E., Pelayo, F.,

Ortega, J., and Rojas, I. (2016). Neural networks: An

overview of early research, current frameworks and

new challenges. Neurocomputing, 214, 242–268.

http://doi.org/10.1016/j.neucom.2016.06.014.

Ramesh, A. N., Kambhampati, C., Monson, J. R. T., and

Drew, P. J. (2004). Artificial intelligence in medicine.

Annals of the Royal College of Surgeons of England,

86(5), 334–338. http://doi.org/10.1308/147870804290.

Remzi, M., Anagnostou, T., Ravery, V., Zlotta, A.,

Stephan, C., and Marberger, M. (2003). An artificial

neural network to predict the outcome of repeat prostate

biopsies. Urology, 62(3), 456–460. http://doi.org/

10.1016/s0090-4295(03)00409-6.

Richards, J. A. (2006). Remote Sensing Digital Image

Analysis: An Introduction. Berlin: Springer-Verlag.

Sargent, D. J. (2001). Comparison of artificial neural

networks with other statistical approaches. Cancer,

91(S8), 1636–1642. http://doi.org/10.1002/1097-0142

(20010415)91:8 <1636::AID-CNCR1176>3.0.CO;2-D

Sazli, M. H. (2006). A brief review of feed-forward neural

networks. Communications Faculty of Sciences

University of Ankara, 50(1), 11–17. http://doi.org/

10.1501/0003168.

Siang, J. J. (2009). Jaringan Syaraf Tiruan and

Pemrogramannya Menggunakan Matlab. Yogyakarta:

ANDI.

Song, J. H., Venkatesh, S. S., Conant, E. A., Arger, P. H.,

and Sehgal, C. M. (2005). Comparative analysis of

logistic regression and artificial neural network for

computer-aided diagnosis of breast masses. Academic

Radiology, 12(4), 487–495. http://doi.org/10.1016/

j.acra.2004.12.016.

Suliman, A., and Zhang, Y. (2015). A Review on Back-

Propagation Neural Networks in the Application of

Remote Sensing Image Classification. Journal of Earth

Science and Engineering, 5

, 52–65. http://doi.org/

10.17265/2159-581X/2015.01.004.

Sun, Y., Peng, Y., Chen, Y., and Shukla, A. J. (2003).

Application of artificial neural networks in the design

of controlled release drug delivery systems. Advanced

Drug Delivery Reviews, 55(9), 1201–1215.

http://doi.org/10.1016/S0169-409X(03)00119-4.

Terrin, N., Schmid, C. H., Griffith, J. L., D'Agostino, R. B.,

and Selker, H. P. (2003). External validity of predictive

models: A comparison of logistic regression,

The 4th ICE on IMERI 2019 - The annual International Conference and Exhibition on Indonesian Medical Education and Research Institute

classification trees, and neural networks. Journal of

Clinical Epidemiology, 56(8), 721–729.

http://doi.org/10.1016/S0895-4356(03)00120-3.

Toshniwal, M. (2005). An optimized approach to

application of neural networks to classification of

multispectral, remote sensing data. In Proceedings.

2005 IEEE Networking, Sensing and Control, 2005.

IEEE.

Tosun, E., Aydin, K., and Bilgili, M. (2016). Comparison

of linear regression and artificial neural network model

of a diesel engine fueled with biodiesel-alcohol

mixtures. Alexandria Engineering Journal, 55(4),

3081–3089. http://doi.org/10.1016/j.aej.2016.08.011.

Ture, M., Kurt, I., Turhan Kürüm, A., and Ozdamar, K.

(2005). Comparing classification techniques for

predicting essential hypertension. Expert Systems with

Applications, 29(3), 583–588. http://doi.org/10.1016/

j.eswa.2005.04.014.

Ul-Saufie, A. Z., Yahya, A. S., Ramli, N. A., and Hamid,

H. A. (2011). Comparison Between Multiple Linear

Regression And Feed forward Back propagation Neural

Network Models For Predicting PM10 Concentration

Level Based On Gaseous And Meteorological

Parameters. International Journal of Applied Science

and Technology, 1(4), 42–49.

Wang, T.-S., Chen, L., Tan, C.-H., Yeh, H.-C., and Tsai,

Y.-C. (2009). BPN for Land Cover Classification by

Using Remotely Sensed Data. In 2009 Fifth

International Conference on Natural Computation (pp.

535–539). IEEE.

Wilson, E. C., Goodman, P. H., and Harris, F. C. (2001).

Implementation of a Biologically Realistic Parallel

Neocortical-Neural Network Simulator. In Proceedings

of the Tenth SIAM on Conference on Parallel

Processing for Scientific Computing.

Artiﬁcial Neural Network and Its Application in Medical Disease Prediction: Review Article