CONSTRUCTING MODELS FOR EARLY CHRONIC KIDNEY
DISEASE DETECTION AND RISK ESTIMATION
Ruey Kei Chiu, Fu Jen Catholic University, Taiwan
rkchiu@mail.fju.edu.tw
Renee Yu-Jing Chen, Fu Jen Catholic University, Taiwan
497746077@mail.fju.edu.tw
ABSTRACT
In this paper, we aim to develop a best-fitting neural network model which can be employed
to detect various severity levels of CKD based on the two different sets of influence factors as
the inputs for model learning and developing. Firstly, the influence factors of CKD which
most physicians employ in the computational formula are used as the inputs to three
preselected fundament neural network models to detect (classify) various levels of chronic
kidney disease. Secondly, the potential key influence factors of CKD which are identified by
reviewing the CKD relevant literatures as well as conducting expert interviews are used as
the inputs to three models as well. Three fundamental neural network models selected for
detection and comparison include back-propagation network (BPN), generalized feedforward
neural networks (GRNN), and modular neural network (MNN). Furthermore, three hybrid
neural network models which are modeled by embedding genetic algorithm (GA) into each
respective fundamental neural factor are employed for conducting CKD detection and
comparison. The comparison among network models are based the CKD detection
performance measured by the computation accuracy, sensitivity, and specificity in model
training and testing.
The results of model experiment show that almost all models employed in experiment may
gain near 100% accuracy in CKD detection in the training stage regardless of which sets of
influence factors used in model training. However, if the model testing is further conducted, it
is found that the network models with the inputs of influence factors of CKD used by
physicians employed in the computational formula always shows better detection
performance in all three aspects of measures. The BPN gain the highest 94.75% accuracy
measures in the testing stage among three fundamental neural network models while GFNN
gain only 86.63% in accuracy measure which is the lowest performance in three models.
Through further observations from experiment results, it is found the hybrid network model of
GFNN plus GA embedded may significantly enhance detection performance in all three
measures from its fundamental model in the testing although the the detection performance
are degraded in BPN and MNN.
We conclude that BPN might be the best-fitting factor Among three fundamental neural
network models employed in detecting CKD while the models with GA embedded in BPN and
GFNN respectively might be the best-fitting hybrid models for CKD detection.
Keywords: Chronic Kidney Disease, Disease Detection, Risk Estimation, Artificial Neural
Networks, Genetic Algorithms.
INTRODUCTION
As people’s life level become more and more modernized and life span become longer in our
society, chronic kidney disease (CKD) becomes more common which may result in
developing different levels of ill-function and damage of patient kidneys. Once a person
gets CKD, he/she will suffer from the disease which may decrease his/her working ability as
well as live quality. It is also very likely to develop other chronic diseases like high blood
pressure, anemia (low blood count), weak bones and cause poor nutritional health and nerve
damage. In the meantime, kidney disease increases patient risk of contracting heart and blood
relevant diseases. Chronic kidney disease may be caused by other chronic disease such as
diabetes, high blood pressure and other disorders. High CKD risk groups include those who
have diabetes, hypertension, and family history of kidney disease. Early detection and
treatment can often keep chronic kidney disease from getting worse. When kidney disease
progresses, it may eventually lead to kidney failure which even requires dialysis or even
proceeds to kidney transplant to maintain patient life.
According to the statistical data announced by the Department of Health of Taiwan’s
government in 2010 (DOH, 2010), the mortality caused by kidney disease has been ranked in
the 10th place in all causes of death in Taiwan and thousands of others are at increased risk.
The mortality caused from kidney disease is estimated as 12.5 in every 100,000 peoples. As a
result, it costs as high as 35 percent of health insurance budget to treat the chronic kidney
disease (CKD) patients with the age over 65 years old and end stage kidney disease patient in
all ages. It occupies a huge amount of expenditures in national insurance budget.
Regarding to the measurement of serious levels of CKD, presently glomerular filtration rate
(GFR) is the most commonly measuring indicators used in health institutions to estimate
kidney health function. The physician in the health institution can calculate GFR from
patient’s blood creatinine, age, race, gender and other factors depending upon the type of
formal-recognized computation formulas (Firman, 2009; NKF, 2002; Levey, 2005) is
employed. The GFR may indicate how health of a patent’s kidney and can be also taken to
determine whether a people has kidney disease. The physician may also depend upon this
indicator to determine the stage of kidney disease of a patient and help him to set a treatment
plan for the patient (Firman, 2009; Levey et al., 2005).
The objectives of this paper aims to develop decision models by applying artificial neural
networks and fuzzy expert system in artificial intelligence which attempts to provide
physicians an alternative method to detect chronic kidney diseases and estimate the risk
levels of a patient in early CKD stage. This study of this paper was cooperatively carried out
with a district teaching hospital in New Taipei City located in the northern district of Taiwan.
In the first stage of this paper, we aim to develop a best-fitting neural network model which
can be employed to identify various levels of CKD based on two different sets of influence
factors as the inputs for model learning and developing. To reach this develop objective,
firstly the set of influence factors of CKD which most physicians employ in the
computational formula to diagnose CDK are determined for being used as the inputs to three
preselected fundament neural network models. Next, another set of potential key influence
factors of CKD which are identified by reviewing the CKD relevant literatures as well as
conducting expert interviews are used as the inputs to three models as well. Three
fundamental neural network models selected for detection and comparison include
back-propagation network (BPN), generalized feedforward neural networks (GRNN), and
modular neural network (MNN). Furthermore, three hybrid neural network models which are
developed by embedding genetic algorithm (GA) into each respective fundamental neural
factor Are employed for conducting CKD detection and comparison. The comparison among
network models is based on the CKD detection performance measured by the measurement
of factor Accuracy, sensitivity, and specificity respectively in model training and testing. The
input data for the training, testing, and experiment of ANN models is collected from the
people’s health examination which is periodically carried out by collaborative hospital.
METHODS AND LITERATURES REVIEW
In this section, firstly the different methods of measuring CKD followed by two sets of
influence factors used as the inputs for network modeling are reviewed. Next, the
investigation of the criteria of CKD classification (i.e., detection) and risk evaluation of CKD
is conducted. The three different neural network models employed for detection and
comparison in this paper are surveyed and briefly introduced as well.
The Review of the Methods of Measuring Chronic Kidney Disease
The GFR is a most-commonly method used to measure kidney health function. It refers to the
water filterability of glomerular of people’s kidney. The normal value of measurement ought
to be between 90 and 120 ml/min/1.73m2. There are three common computation methods of
GFR, which are (1) Removing rate of 24-hour urine creatinine (i.e., Creatinine Clearance
Rate, Ccr) ; (2) Cockcroft-Gault formula (i.e., known as C-G formula) ; (3) Modification of
Diet in Renal Disease formula (i.e., known as MDRD) (Firman, 2009). The CKD is
categorized into five stages by making using of GFR to measure kidney function. Although
the course of the change from stage one to five may usually last for years, it sometimes may
enter into fifth stage pretty soon resulted in the necessity of dialysis or kidney transplant
(NKF, 2002; Levey et al., 2005).
Again, according to the guideline developed by the National Kidney Foundation’s Kidney
Disease Outcomes Quality Initiative (KDOQI) (NKF, 2002), an independent not-for-profit
foundation governed by an international Board of Directors in 2002 providing a new
conceptual framework for a diagnosis of CKD independent of cause and developing a
classification scheme of kidney disease severity based on the level of glomerular filtration
rate (GFR). The new system represented a significant conceptual change, since kidney
disease historically had been categorized mainly by cause. The diagnosis of CKD relies on
markers of kidney damage and/or a reduction in GFR. Stages 1 and 2 define conditions of
kidney damage in the presence of a GFR of at least 90 ml/min/1.73 m2 or 60 to 89
ml/min/1.73 m2, respectively, and stages 3 to 5 define conditions of moderately and severely
reduced GFR irrespective of markers of kidney damage. The summarized of this guideline is
shown in Table 1. The common definition of CKD has facilitated comparisons between
studies. Thus, this new diagnostic classification of CKD has likely been one of the most
profound conceptual developments in the history of nephrology. Nevertheless, there are
limitations to this classification system, which is by its nature simple and necessarily arbitrary
in terms of specifying the thresholds for definition and different stages. When the
classification system was developed in 2002 (NKF, 2002; Levey et al., 2005), the evidence
base used for the development of this guideline was much smaller than the CKD evidence
base today. Therefore, this guideline is constantly revised from then on by Kidney Disease
Outcomes Quality Initiative (KDOQI) with the modification and endorsement by Kidney
Disease: Improving Global Outcomes (KDIGO). The currently available methods to estimate
GFR and ascertain kidney damage are evolving (Levey et al., 2005).
Table 1: Classification of CKD as Defined by KDOQI and Modified and Endorsed by
KDIGO.
Stage Description
Classification by Severity
By GFR (ml/min/1.73m2)
1
Kidney damage with normal or GFR increasing
GFR >=90
2
Kidney damage with mild decreasing in GFR
GFR of 60-89
3
Moderate with decreasing in GFR
GFR of 30-59
4
Severe with decreasing in GFR
GFR of 15-29
5
Kidney failure
GFR <= 15 (or dialysis)
One of most recognized versions of the classification and risk evaluation of CKD was
released by National Kidney Foundation (NKF) of US written by Firman in 2009 (Firman,
2009). Its severity classification and risk evaluation criteria are shown in Table 2 and 3,
respectively. It is properly the most referable guideline regarding the classification and risk
evaluation of CKD.
Table 2: Classification and Risk Evaluation of Chronic Kidney Disease.
Stage
Description†
GFR (ml per minute Action plans
per 1.73 m2)
-
At increased risk
> 60 (with risk factors
Screening, reduction of risk factors for chronic
for chronic kidney
for chronic kidney
kidney disease
disease
disease)
Kidney damage
> 90
1
Diagnosis and treatment, treatment of comorbid
with normal or
conditions, interventions to slow disease
elevated GFR
progression, reduction of risk factors for
cardiovascular disease
2
Kidney damage
60 to 89
Estimation of disease progression
30 to 59
Evaluation and treatment of disease
with mildly
decreased GFR
3
Moderately
decreased GFR
4
Severely decreased
complications
15 to 29
GFR
5
Kidney failure
Preparation for kidney replacement therapy
(dialysis, transplantation)
< 15 (or dialysis)
Kidney replacement therapy if uremia is
present
Table 3: Risk Factors for Chronic Kidney Disease and Definitions
Type
Definition
Examples
Susceptibility
Factors that increase
Older age, family history of chronic kidney disease,
factors
susceptibility to kidney damage
reduction in kidney mass, low birth weight, U.S.
racial or ethnic minority status, low income or
educational level
Initiation
Factors that directly initiate
Diabetes mellitus, high blood pressure, autoimmune
factors
kidney damage
diseases, systemic infections, urinary tract
infections, urinary stones, obstruction of lower
urinary tract, drug toxicity
Progression
Factors that cause worsening
Higher level of proteinuria, higher blood pressure
factors
kidney damage and faster decline
level, poor glycemic control in diabetes, smoking
in kidney function after kidney
damage has started
End-stage
Factors that increase morbidity
Lower dialysis dose (Kt/V)*, temporary vascular
factors
and mortality in kidney failure
access, anemia, low serum albumin level, late
referral for dialysis
*-In Kt/V (accepted nomenclature for dialysis dose), "K" represents urea clearance, "t"
represents time, and "V" represents volume of distribution for urea.
In Taiwan, the Taiwan Society of Nephrology (TSN) also presented the self-detecting method
of kidney for the public. The MDRD formula (Levey et al., 2000; Levey et al., 2005), which
is recognized as a more accurate with mostly adopted method by kidney physicians to
estimate glomerular filtration rate (GFR) from serum creatinine, is adopted for the use in
diagnosis. Therefore, in this paper of study we also take MDRD to calculate the GFR. The
method of calculation formula is shown in formula 1. The input data for the GFR calculation
of each individual patient case in health examination are provided by the collaborative
hospital of this research. We also use the calculated results as the desired (targeted) value to
train our neural network models in experiment.
GFR = 186 × Creatinine−1.154 × age−0.203 (ml/min/1.73m2 )…….(1)
Note: For female the result should be multiplied by a factor of 0.742。
Snyder and Pendergraph (2005) indicated that the MDRD was suitable to be used to the
group of diabetes mellitus, older age (age above 51), patient with kidney replacement therapy
(dialysis, transplantation). It also takes less number of factors than other two popular methods,
which are Creatinine Clearance Rate computation and Cockcroft-Gault formula. It becomes
easier and simple but still maintains the result effectiveness for calculating GFR. Hence, this
paper adopts MDRD to conduct the calculation of GFR directly from the data collection of
creatinine, age, and sex. The three factors, which are ‘creatinine’, ‘age’, and ‘sex’, considered
in MDRD calculation also majorly compose the first set of input influence factors for
modeling neural network models.
Artificial Neural Network
An artificial neural network (ANN), usually called neural network (NN), is a mathematical
model or computational model that is inspired by the structure and/or functional aspects of
biological neural networks. Kriesel (2011) indicates neural networks are a bio-inspired
mechanism of data processing that enables computers to learn technically similar to
human-being brain. The generic structure of ANN is shown in Figure 1 (Nagnevitsky, 2005).
A neural network consists of an interconnected group of artificial neurons, and it processes
information using a connectionist approach to computation. In most cases an ANN is an
adaptive system that changes its structure based on external (input) or internal information
that flows through the network during the learning phase. Modern neural networks are
non-linear statistical data modeling tools. They are usually used to model complex
relationships between inputs and outputs or to find patterns in data
(http://en.wikipedia.org/wiki/Artificial_neural_network). It is properly the most prestigious
and adoptable model in all application models in the field of artificial intelligence.
Figure 1: A Generic Structure of ANN.
Input
layer
First
hidden
layer
Second
hidden
layer
Output
layer
There are many types of neural network models derived from the generic structure of ANN.
Three fundamental neural network models employed for the experiment in this paper are
briefly introduced in next sections
Back-Propagation Neural Network
The back-propagation neural network is aslo known as a "feed-forward back-propagation
network". It is a supervising learning multi-layer perceptron (MLP) neural network model
(Negnevitsky, 2005). A training set of input patterns is presented to the network. The network
computes its output pattern, and if there is an error - or in other words a difference between
actual and desired output patterns - the weights are adjusted to reduce this error. Since the
real uniqueness or 'intelligence' of the network exists in the values of the weights between
neurons, we need a method of adjusting the weights to solve a particular problem. It learns by
example, that is, we must provide a learning set (or called as training set) that consists of
some input examples and the known-correct output (desired output) for each case. So, we use
these input-output examples to show the network what type of behavior is expected, and the
back-propagation learning algorithm allows the network to adapt. Hence, this neural
network is called as back-propagation neural network (BPNN). A three-layer architecture of
BPNN is illustrated in Figure 2 (Negnevitsky, 2005).
Input signals
x1
x2
1
2
2
xi
i
1
y1
2
y2
k
yk
l
yl
1
wij
j
wjk
m
xn
n
Input
layer
Hidden
layer
Output
layer
Error signals
Figure 2: Three-layer Back Propagation Neural Network.
This model is properly the most popular and applicable model in artificial neural network
used in academic research and the development practical decision support applications in
terms of forecasting, classification, regression, data mining, and so on.
Generalized Feedforward Neural Network
The feedforward neural network was the first and arguably most simple type of artificial
neural network devised. In this network the information moves in only forward direction
which is simply from the input neuron data goes through the hidden neuron and to the output
nourons. There are no cycles or loops in the network where the data flow from input to output
units is strictly feedforward. The data processing can extend over multiple layers of neurons,
but no feedback connections are present. Its generalization model, which is konwn as
generalized feedforward neural networks (GFNN). It has the same architecture as
back-propagation neural network and also adopts the same learning methods to train the
network. The differences are that GFNN adopts generalized distributed-flow neurons in its
hidden layers and take perceptron as its inhibitory neuron. Arulampalam and Bouzerdoum
(2000) mentioned that the GFNN architecture used as the basic computing unit is a
generalized shunting neuron (GSN) model which includes as special cases the perceptron and
the shunting inhibitory neuron. GSNs were capable of forming complex, nonlinear decision
boundaries. This allows the GFNN architecture to easily learn some complex pattern
classification problems. Arulampalam and Bouzerdoum (2003) applied GFNNs to several
benchmark classification problems, and their performance was compared to the performances
of multilayer perceptrons such as a back-propagation neural network. Experimental results
show that a single GSN can outperform multilayer perceptron networks. An exemplified
architecture of GFNN is illustrated as Figure 3.
Generalized
Shunting
w00, c00 Neurons
Input Neurons
x0
Output Neurons
(Perceptorns)
wk0
‧
‧
‧
wm0, cm0
‧
‧
‧
o0
‧
‧
‧
xm
wjm, cjm
‧
‧
‧
‧
‧
‧
ok
wkj
xi
wji, cji
Excitatory Synapses
Inhibitory Synapses
Figure 3: The Generic Architecture of Generalized Feedforward Neural Networks.
Modular Neural Network
The modular neural network (MNN) was initially presented by Jacobs, et al. in 1991 (Jacobs
et al. 1991). The architecture of this network is shown in Figure 4. It is characterized by a
series of independent neural networks moderated by some intermediary. Each independent
neural network serves as a module like an expert and operates on separate inputs to
accomplish some subtask of the task the network hopes to perform. The intermediary takes
the outputs of each module and processes them to produce the output of the network as a
whole (Azam, 2000). The intermediary only accepts the modules’ outputs—it does not
respond to, nor otherwise signal, the modules. As well, the modules do not interact with each
other. One of the major benefits of a modular neural network is the ability to reduce a large,
unwieldy neural network to smaller, more manageable components. There are some tasks it
appears are for practical purposes intractable for a single neural network as its size increases.
The following are benefits of using a modular neural network over a single all-encompassing
neural network.
Expert 1
×
x
×
Expert 2
Σ
y
×
Expert k
g3
g2
g1
Gate
Figure 4: The Generic Architecture of Modular Neural Networks
(Source: http://en.wikipedia.org/wiki/Modular_neural_networks).
RESEARCH MATERIALS AND METHODS
Based on the literature reviews and expert interviews, this research takes three neural network
models, back propagation neural network, feedforward neural network, and modular neural
network, respectively to generate the detection model for chronic kidney disease. The input
data set for neural networks training and testing are collected from the population data of
health examination provided by collaborative hospital. The key influence factors input for the
each factor Are determined and verified from professional kidney experts and physicians. The
set of influence factors derived from GFR computation formulas such as MDRD is shown in
Table 4 and another set of key influence factors selected and determined in this paper as the
inputs of Neural Networks is shown in Table 5, respectively. The classification performance
of two sets input for model training and testing are compared in model experiment.
Table 4: The Influence Factors Derived from Formula Computation for GFR.
Factors
Description
Mean
Maximum
Minimum
Sex
0→Female;1→Male
---
---
---
Age
Integer
77.38
93
67
Creatinine
Real
0.88
4.28
0.37
Weight
Real
60.719
98
32
GFR Levels
0→Negative;1, 2, 3,4,5→Positive
---
---
---
measured in levels
Table 5: The Influence Factors Investigated Determined by This Paper.
Factors
Description
Mean Max. Min.
Creatinine
Real
0.88
4.28
0.37
Glucose (GLU)
Real
109.63
271
82
Systolic Pressure (SP)
Real
135.78
195
88
Protein in Urine (UP)
0→Negative
---
---
---
or called proteinuria
1→Trace; 2→+
---
---
---
3→++; 4→+++
Hematuria
Blood Urea Nitrogen (BUN)
Real
15.80
80.1
5.4
GFR-Level
0→Negative;1, 2, 3,4,5→Positive
---
---
---
measured in levels
The Learning Data and Preprocessing
The input learning data set for neural networks are collected and sampled from the cases of
the elderly people with age over 65 years old of health examination provided by collaborative
hospital in past few years. The areas of sampled data cover the hospital neighboring cities and
counties located in northern part of Taiwan. There are 1161 heath examination case records
are sampled. Before they are input for training and testing network models, a data
preprocessing is processed to remove duplicate, correct the error, inconsistency and missing
fields, numerification of class field in each case record.
By this process, we can ensure the accuracy, completeness and integrity of the input data.
After data preprocessing, only 945 sampled cases are left as the use of learning data set.
Meanwhile, initially there are 21 fields contained in each source of case record. Only 6 fields
specified in Table 5 are selected and maintained for each case.
Model Learning and Experiments
We take NBuilder of neural solution, a well-known neural network modeling tool, to train
and test our neural network model through a series of simulations in experiment. In
consideration of the tool restriction for building the hybrid model of neural network and
genetic algorithm which is also studied in this research, we only take the most accurate 430
cases from 945 sampled cases left after preprocessing in learning process. By this selection,
we can ensure there are equal numbers of cases used in the leaning process of each factor
Building. As it is shown in Table 6, among 430 cases, 300 cases are selected for training, 100
records are used for testing, and 30 cases for cross verification. The desired output of
diagnosis is identified by the professional physicians of chronic kidney disease from the
collaborative hospital with the aid of MDRD calculation formula.
Table 6: Data Distribution for Model Training, Testing, and Cross Verification.
Sampled Data
Record #
Desired Output #
Training Data
300
101 Negative
199 positive in various
levels
Testing Data
100
34 negative
66 positive
Cross Verification
30
10 negative
20 positive
EXPERIMENTS AND COMPARISONS OF NEURAL NETWORK MODELS
The experiment for a specific network model can be viewed as a series of modeling and
simulation through leaning process. We attempt to pursue a best model through the selection
of different combination of modeling parameters for conducting learning in different network
models. The parameters combination shown in different model experiments have been
filtered from a series of learning process.
The Experiment of Back-propagation Neural Network
Table 7 shows the parameters set for the experiment of back-propagation neural network.
Table 7: Parameters Selected for Back-propagation Neural Network.
Parameters
Value Setting
# of Hiding Layers
1、2、3
# of Neurons in Hiding Layer
5、8、12
Learning Rules
Step、Momentum、ConjugateGradient、
LevenbergMarquar、Quickprop、DeltaBarDelta
Transformation Functions
TanhAxon、SigmoidAxon、LinearTanhAxon、
LinearSigmoidAxon、SoftMaxAxon、BiasAxon、LinearAxon、
Axon
Weight Update Methods
Batch
Criterion for Termination
MSEmin:0.0001; Epochsmax:3000
In this paper, we also take the hybrid models of combining each neural network factor And
genetic algorithm (GA) in experiment and comparison. The model parameters selected for
genetic algorithm are shown in Table 8. GA will be taken to combine with back-propagation
neural network, feedforward neural network, and modular neural network, respectively in
model experiment and comparison as well.
Table 8: The Parameters Selected for Genetic Algorithm.
Network Parameters
Value Setting
Method of Selection
Roulette
Method of Crossover
One Point
Rate of Crossover
0.9
Rate of Mutation
0.01
Termination Criterion
MSEmin:0.0001
Epochsmax:100
Population: 20
Generations：50
Through a series of simulation by taking the different sets of parameter combination selected
from Table 7, we gain that the best model of backpropagation neural network in terms of their
respective network parameters with respect to two models, which are the factor By adopting
the key factors used in computation formula (called as Factor A hereafter) and the factor By
adopting the key factors selected in this paper (called as Factor B hereafter), respectively, are
listed in Table 9. The classification accuracy, sensitivity, and specificity with respect to factor
A and B, respectively is shown in Table 10.
Table 9: The Best Models with Backpropagation Network Parameters Setting.
Network Parameters
By Adopting the
Key Factors Used
in Computation
Formula (Factor
A)
By Adopting the
Key Factors
Selected in This
Paper (Factor B)
Learning Rule
LevenbergMarquar
LevenbergMarquar
Transformation Functions
TanhAxon
SigmoidAxon
# of Hiding Layer
2
2
# of Neurons in Hiding Layer
5
8
Criterion for Termination
MSEmin:0.0001
MSEmin:0.0001
Epochsmax:3000
Epochsmax:3000
Batch
Batch
Methods for Weight Update
Table 10: The Experiment Results from Backpropagation Neural Network Models.
BPN
Factor A
Factor B
Pure BPN
Training
Test
Training
Test
Accuracy of
100%
94.75%
100%
72.42%
Sensitivity
100%
97.06%
100%
70.59%
Specificity
100%
92.42%
100%
74.24%
BPN+GA
Test
Test
Accuracy of
91.71%
77.68%
Sensitivity
97.06%
81.82%
Specificity
86.36%
73.53%
Classification
Classification
In Table 10, “accuracy” is used to measure the classification accuracy with the proportion of
the sum of the number of true positives and the number of the true negatives which are
correctly identified. “Sensitivity” is used to measure the proportion of actual positives
which are correctly identified as such (e.g. the percentage of sick people who are correctly
identified as having the condition). “Specificity” is used to measure the proportion of
negatives which are correctly identified (e.g. the percentage of healthy people who are
correctly identified as not having the condition). These two measures are closely related to
the concepts of type I and Type II. From the results shown in Table 10, our experiment may
gain very good results with BPN in model training stage but it shows a significant drop
measured in model testing stage both with BPN and BPN plus GA. By the results shown in
Table 10, BPN may obtain better results in accuracy measure, but GPN plus GA gain better
result in sensitivity measure. The result also shows the model with the adoption of the key
factors used in computation formula gains better results in all three indicators measures.
These results tell us that a hybrid model with the combination of BPN and GA does not
improve the model performance in accuracy measure but it is helpful to improve sensitivity
measure.
The Experiment of Generalized Feedforward Neural Network
Table 11 shows the parameters set for the experiment of generalized feedforward neural
network (GFNN). The classification accuracy, sensitivity, and specificity with respect to
factor A and B, respectively is shown in Table 12. As the results shown in Table 12, pure
GFNN obtain a perfect classification percentage in three measures including accuracy,
sensitivity, and specificity in training stage, but GFNN plus GA apparently may gain better
results in testing. However, they also show the results gained from GFNN are not so good in
all three measures as those gained from BPN and BPN plus GA, respectively.
Table 11: The Best Models with Generalized
Feedforward Neural Network Parameters Setting.
Network
Parameters
Factor A
Factor B
Learning Rule
LevenbergMarquar
LevenbergMarquar
Transformation
SigmoidAxon
TanhAxon
2
2
8
8
Criterion for
MSEmin:0.0001
MSEmin:0.0001
Termination
Epochsmax:3000
Epochsmax:3000
Methods for
Batch
Batch
Functions
# of Hiding
Layer
# of Neurons
in Hiding
Layer
Weight Update
Table 12: The Experiment Results from Generalized Feedforward Neural Network Models.
GFNN
Factor A
Factor B
Pure GFNN
Training
Test
Training
Test
Accuracy of Classification
100%
86.63%
100%
71.03%
Sensitivity
100%
82.35%
100%
61.76%
Specificity
100%
90.91%
100%
80.30%
GFNN+GA
Test
Test
Accuracy of Classification
91.09%
78.39%
Sensitivity
88.24%
76.47%
Specificity
93.94%
80.30%
The Experiment of Modular Neural Network
Table 13 shows the parameters set for the experiment of modular neural network (MNN). The
classification accuracy, sensitivity, and specificity with respect to factor A and B, respectively
is shown in Table 14. As the results shown in Table 13, pure MNN same as prior two
models which may obtain a perfect classification percentage in three measures in training
stage, but MNN plus GA apparently may gain worse results in testing stage which are
different from the results shown in prior two models.
Table 13: The Best Models with Modular Neural Network Parameters Setting.
Network
Parameters
Factor A
Factor B
Learning Rule
LevenbergMarquar
LevenbergMarquar
Transformation
SigmoidAxon
SigmoidAxon
# of Hiding Layer
2
2
# of Neurons in
4
8
Criterion for
MSEmin:0.0001
MSEmin:0.0001
Termination
Epochsmax:3000
Epochsmax:3000
Methods for
Batch
Batch
Functions
Hiding Layer
Weight Update
Table 14: The Experiment Results from Modular Neural Network Models.
MNN
Factor A
Factor B
Pure
MNN
Training Test
Training Test
Accuracy of
100%
93.23%
100%
70.99%
Sensitivity
100%
97.06%
100%
64.71%
Specificity
100%
89.39%
100%
77.27%
MNN+GA
Test
Test
Accuracy of
88.82%
78.34%
Sensitivity
89.39%
79.41%
Specificity
88.24%
77.27%
Classification
Classification
CONCLUSIONS
The performances of three neural network models developed for comparison in this paper in
terms of detection (i.e., classification) accuracy measure are summarized and shown in Table
15 and 16, respectively. Table 15 shows the detection accuracy in three pure (i.e., fundamental)
neural network models while Table 16 shows the detection accuracy with GA embedded in
three respective models. We found BPNN may obtain best accuracy regardless of in factor A
or in factor B which are measured with 94.75% and 72.42% accuracy, respectively in model
testing. However, factor A, which is the test by adopting the key factors used in computation
formula, may gain much better performance than factor B, which is the test by adopting the
key factors selected in this paper. The same result is gained with the GA embedded in each
fundamental model in experiment. As the results observed from Table 16, both BPN plus GA
and GFNN plus GA gain close results in accuracy measure which is much better than MNN
plus GA.
Table 15: Detection Accuracy in Three Pure Neural Network Models.
Network
Models
By Adopting the
Key Factors Used
in Computation
Formula (Factor A)
By Adopting the
Key Factors
Selected in This
Paper (Factor B)
BPN
94.75%
72.42%
GFNN
86.63%
71.03%
MNN
93.23%
70.99%
Table 16: Detection Accuracy in Three Pure Neural Network Models Plus GA.
Network
Models
Factor A
Factor B
BPN+GA
91.71%
77.68%
GFNN+GA
91.09%
78.39%
MNN+GA
88.82%
78.34%
By further observation from the results of model experiment shown in section 4, it is found
that almost all models employed in experiment may gain near 100% accuracy in CKD
detection in the training stage regardless of which sets of influence factors used in model
training. However, if further model testing is conducted, it is found that the network models
with the inputs of influence factors of CKD used by physicians employed in the
computational formula always shows better detection performance in all three aspects of
measure including accuracy, sensitivity, or specificity measures. The BPN gain the highest
94.75% (shown in Table 10 and 15) accuracy measure in the testing stage among three
fundamental neural network models while GFNN gains only 86.63% (shown in Table 12 and
16) in accuracy measure which is the lowest performance in three models. Through further
observations from experiment results, it is found the hybrid network model of GFNN plus GA
embedded may significantly enhance detection performance in all three measures from its
fundamental model in the testing although the reversed enhancement are shown in BPN and
MNN. Again, by the experiment figures shown in section 4, although pure neural network
models plus GA shows degraded performance improvement for factor A in testing stage in
terms of three measures in accuracy, sensitivity, and specificity but it shows significant
improvement for factor B. As a result, we conclude a hybrid model indeed may improve
detection performance. This result may conclude that GA provides no benefit in the yield of
better detection performance with the adoption of the key factors used in computation formula
as the inputs to all models in testing.
Above all, we conclude that BPN might be the best-fitting factor Among three fundamental
neural network models employed in detecting CKD while the models with GA embedded in
BPN and GFNN respectively might be best-fitting hybrid models for CKD detection but only
GFNN plus GA can gain enhancement in detection performance in both sets of influence
factors as inputs. These conclusions should be further verified and compared with other
models selected for experiment in later studies before they can be assured.
We believe that the neural network models respectively developed for CKD detection and risk
level estimation in this paper may equip medical staff with the ability to make precise
diagnosis and treatment to the patient. Meanwhile, public users may take advantage of these
system instruments to conduct self-detection of the risk level of causing CKD in order to take
necessary precaution in advance to avoid any risk of causing CKD or prevent their condition
from getting worse.
REFERENCES
1. Azam F. (2000). Biologically Inspired Modular Neural Networks. Ph.D. dissertation,
Virginia Polytechnic Institute and State University.
2. Arulampalam G., & Bouzerdoum, A. (2003). A Generalized Feedforward Neural Network
Classifier. Proceedings of the International Joint Conference on Neural Networks, 2,
1429-1434.
3. Arulampalam G., & Bouzerdoum A. (2000). Training shunting inhibitory artificial neural
networks as classifiers. Neural Network World, 3(10), 333-350.
4. Arulampalam G., & Bouzerdoum A. (2001). Application of shunting inhibitory artificial
neural networks to medical diagnosis. Proceedings of the Seventh Australian and New
Zealand Intelligent Information Systems Conference (ANZIIS 2001), Perth, Australia,
89-94.
5. Chiu, R.K., Huang C.I., & Chang Y.C (2010). The Study on the Construction of
Intelligence–Based Decision Support System for Diabetes Diagnosis and Risk Evaluation.
Journal of Medical Informatics, Taiwan Association for Medical Informatics, 19(3), 1-24.
6. Department of Health, Taiwan (DOH) (2010). The Statistical Data of the Causes of
Civilian Mortality, Retrieved from http://www.doh.gov.tw/CHT2006/DM/DM2_2.aspx?
now_fod_list_no=11122&class_no=440&level_no=3.
7. Eknoyan, G. (2007). Chronic kidney disease definition and classification: The quest for
refinements. Kidney International, 72, 1183-1185.
8. Kochurani, O.G., Aji, S., & Kaimal, M.R.(2007). A Neuro Fuzzy Decision Tree Model for
Predicting the Risk in Coronary Artery Disease. Proceedings of 22nd IEEE International
Symposium on Intelligent Control, Singapore, 166-171.
9. Firman G. (2009), Definition and Stages of Chronic Kidney Disease (CKD), Released on
Feb. 28, 2009, Retrieved from http://www.medicalcriteria.com/site/index.php?option=com
_content&view=article&id=142%3Anefckd&catid=63%3Anephrology&Itemid=80&lang
=en.
10. Kriesel, D. (2011). A Brief Introduction to Neural Networks, Retrieved from
http://www.dkriesel.com.
11. National Kidney Foundation (NKF) (2002). K/DOQI Clinical Practice Guidelines for
Chronic Kidney Disease: Evaluation, Classification and Stratification. American Journal
of Kidney Disease, 39, S1-S266.
12. Levey, A.S., Greene, T., Kusek, J.W, Beck, G.L., et al. (2000). MDRD Study Group. A
simplified equation to predict glomerular filtration rate from serum creatinine. J Am Soc
Nephrol, 11:155A.
13. Levey, A.S. Eckardt, K.U., & Tsukamoto, Y. (2005). Definition and classification of
chronic kidney disease: A position statement from Kidney Disease: Improving Global
Outcomes (KDIGO). Kidney International, 67, 2089-2100.
14. Snyder, S., & Pendergraph, B. (2005). Detection and Evaluation of Chronic Kidney
Disease. American Family Physician, 72(9), 1723-1732.
15. Jacobs, R., Jordan, MNowlan, S., & Hinton, G. (1991). Adaptive Mixtures of Local
Experts. Neural Computation, 3(1), 79-87.