Diagnosing Chronic Obstructive Pulmonary Disease

with Artiﬁcial Neural Networks using Health Expert Guidelines

Maria Ang

elica de Oliveira Camargo-Brunetto

and Andr

e R. Gonc¸alves

Computer Science Department, State University of Londrina, Londrina, Paran

a, Brazil

School of Electrical and Computer Engineering, University of Campinas, Campinas, S

ao Paulo, Brazil

Keywords:

Decision Support System, Artiﬁcial Neural Networks, COPD, Multilayer Perceptron, Radial Basis Function,

Least Mean Squares.

Abstract:

Chronic Obstructive Pulmonary Disease (COPD) is characterized by airﬂow limitation and the spirometry is

one of the tests that can be used to detect such disease. However there is a great problem related to the different

ways of interpreting the values provided by spirometric devices, regarding different guidelines and reference

values. Artiﬁcial Neural Networks (ANN) can be used to help with tasks of diagnosis as that. This work

presents the modeling and analysis of three ANN models to classify subjects with COPD, based on different

sets of variables: a set of observed measures from spirometry and a set of interpreted values according to the

guideline proposed by the American Thoracic Society. The results shown that it is possible to achieve a good

accuracy in the diagnosis of COPD using ANNs, besides these features set conducted the COPD identiﬁcation

problem to a nearly linearly separable classiﬁcation problem.

1 INTRODUCTION

According to the Global Initiative for Chronic Ob-

structive Lung Disease (GOLD), Chronic Obstructive

Pulmonary Disease (COPD) is a disease state char-

acterized by airﬂow limitation that is not fully re-

versible. In (Group, 2004), it is reported that COPD

is placed just after cardiovascular and neoplastic dis-

eases as a leading cause of death and morbidity. It

also implies a high rate of medical services, as well as

hospitalization, frequently, for a long time (Lenfant,

1998).

Diagnosing and monitoring progression of COPD

is commonly done by spirometry, considered the gold

standard for such tasks. However, differences in the

deﬁnition of COPD and consensus statements make it

difﬁcult diagnosing, as pointed out by (Nathell et al.,

2007). In this paper, the authors reported that the

use of the European Respiratory Society Guidelines

(ERS) has proved to be more effective in detecting pa-

tients with COPD than other guidelines, as the GOLD

COPD-criteria and the NICE COPD-criteria. The au-

thors also show how the COPD diagnosis is highly

dependent on the which guidelines are used for deﬁn-

ing the disease.

Artiﬁcial Neural Networks (ANN) have been

commonly used as classiﬁers or predictors in several

ﬁelds of knowledge, including biomedicine. They

can be applied in clinical medicine, physiological

sign processing and medical image processing. Pas-

sold et al. (Passold et al., 1996) resume the ben-

eﬁts of ANN applied in biomedicine in three main

points: (i) Simulation of human reasoning in diagno-

sis, based on a given data set; (ii) Learning ability in a

self-organized way; and (iii) High performance when

compared to statistical methods. The task of diagnos-

ing diseases, which present some level of difﬁculty for

human consensus, has been successfully solved using

ANN (Zhou and Jiang, 2003), (Wadie et al., 2006),

(Yan et al., 2006), and (Mehrabi et al., 2009).

In this context, the aim of this article is to study

three ANN models to classify subjects with COPD,

based on different sets of variables: a set of observed

measures from spirometry and a set of interpreted val-

ues according to the guideline proposed by the Amer-

ican Thoracic Society, that is commonly included in

the software provided by many spyrometers.

The remaining of the paper is structured as fol-

lows: Section 2 presents related work, Section 3 de-

scribes the ANN modeling. The performance of each

model is analyzed in Section 4. Conclusions and fu-

ture work are presented in Section 5.

207

de Oliveira Camargo-Brunetto M. and R. Gonçalves A..

Diagnosing Chronic Obstructive Pulmonary Disease with Artiﬁcial Neural Networks using Health Expert Guidelines.

DOI: 10.5220/0004234102070214

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2013), pages 207-214

ISBN: 978-989-8565-37-2

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

2 RELATED WORKS

Fontenla-Romero et al. (Fontenla-Romero et al.,

2005) presented a new method to classify sleep ap-

nea as: obstructive, central or mixed. However, this

classiﬁcation method requires to analyze a large num-

ber of variables during a long time. So, the inputs of

the neural network were pre-processed by applying a

discrete wavelet transformation on the samples, aim-

ing to reduce and to ﬁx the number of inputs of the

classiﬁer.

Basically, the authors used wavelets to extract fea-

tures from the recorded signals, that become the input

of a feed-forward ANN in order to do the classiﬁca-

tion. Three learning algorithms were used. Conjugate

Gradient Method was employed to adjust the weights

and the quadratic mean error as the cost function. A

similar algorithm, adds a regularization term to avoid

hyper-training. Finally, a learning algorithm uses a

Bayesian framework and a cross-entropy error func-

tion. The latter showed to be the best choice, with an

accuracy of 83.78%±1.90%.

Er and Temurtas (Er and Temurtas, 2008) present

a comparative study for the realization of the COPD

diagnosis using multilayer neural networks. The

authors applied two different Multilayer Perceptron

(MLP) structures: one with one hidden layer and the

other with two hidden layers. The COPD dataset

was taken from the Diyarbakir Chest Diseases Hos-

pital from Southeast of Turkey. The dataset con-

tains 155 samples, where 55 are COPD and 100 nor-

mal. They analyzed 38 features from laboratory ex-

amination. Accuracy of 93.14% was obtained for an

MLP ANN with one hidden layer and backpropaga-

tion algorithm with momentum constant. Using the

Levenberg-Marquardt learning algorithm, the accu-

racy achieved was 94.46%. Best results were obtained

using MLP with two hidden layers, where the accura-

cies for same learning algorithms were 95.43% and

96.08%.

Mehrabi et al (Mehrabi et al., 2009) used a

MLP and a Radial Basis Function Neural Network

(RBFNN) to classify patients with COPD and Con-

gestive Heart Failure (CHF). They analyzed 266 pa-

tients, being 129 with CHF and 137 with COPD. It

was considered 42 clinical variables. A ten-fold cross

validation was used to assess the generalization of the

classiﬁcation models and the results obtained were:

MLP sensitivity of 83.9% and speciﬁcity of 86%, with

AUC (Area under the ROC curve) of 0.889 ±0.002;

RBFNN sensitivy of 81.8% and speciﬁcity of 88.4%,

with AUC of 0.924 ±0.01.

Er et al (Er et al., 2010) analyzed several classi-

ﬁers in order to classify patients among the follow-

ing chest diseases: COPD, Pneumonia, Asthma, Tu-

berculosis and Lung Cancer Diseases. The following

models were used: MLP (with one and two hidden

layers), without and with momentum; Probabilistic

Neural Network, Generalized Regression Neural Net-

work, Learning Vector Quantization and RBFNN. A

sample of 357 patients was analyzed, where 71 with

COPD, 50 with Tuberculosis, 60 with Pneumonia, 44

with asthma, 32 with lung cancer, and 100 normal.

It was used 38 variables as input for all ANN mod-

els. For the COPD diagnosis, the best results were

obtained with MLP with two hidden layers and Prob-

abilistic Neural Network, with accuracy of 88.73%.

The results achieved with the application of ANN

as classiﬁers of diseases have encouraged researchers

study and to apply such models to help with tasks of

COPD diagnosis. COPD is a chronic disease that re-

quires constant care as well as knowledge of its sever-

ity, that is normally done by spirometry. Normally,

the devices that perform such exam have softwares

that provide different reference values and guidelines

to identify the degree of normality of the subject.

Since COPD diagnosing is highly dependent on the

guideline used for detecting it, ANN could be an al-

ternative to help health care professionals to classify

subjects with COPD.

3 ANN MODELING FOR COPD

IDENTIFICATION

This research considered three ANN models based

on supervised learning: a linear ANN using the

LMS algorithm, a Multilayer Perceptron with back-

propagation algorithm and a Radial Basis Function

Neural Network. A detailed description of each one

is presented in Section 3.3. The physiologic measures

used for this study were obtained from pulmonary

function test and performed at the Physiotherapy Pul-

monary Laboratory at State University of Londrina,

Brazil.

3.1 Obtaining the Physiological

Measures

The pulmonary function test is composed of several

exams that provide information about the pulmonary

capacity of a subject. It is composed of spirome-

try, ventilometry, and measures of the inspiratory and

expiratory pressures. The spirometry measures the

volume, capacity, and pulmonary ﬂow, from respi-

ratory maneuver and compares them with reference

values of normality to the evaluated population. For

HEALTHINF2013-InternationalConferenceonHealthInformatics

208

all volumes and capacities measured by the spirom-

etry, there are reference tables, which predict values

are based on regression equations. Several guidelines

and normality standards have been proposed by scien-

tiﬁc societies e.g. ATS (American Thoracic Society),

ERS (European Respiratory Society), ALATS (Amer-

ican Latin Thoracic Society), and others. Nowadays,

the spirometry has been used in the respiratory phys-

iotherapy as a complementary exam and has been

showed very helpful to physiodiagnosis and to plan

a therapeutic program. The spirometer used in this

work provides a few options of reference values, and

the ATS pattern was selected by the healthcare pro-

fessional that did the exam (Azeredo, 2002).

The obtained values by spirometry provide rich in-

formation about the pulmonary function, helping to

identify and to qualify the severity of several venti-

latory disturbs that are observed by changes in the

spirometric values. It must be remembered that the

interpretation of the results is highly dependent on

the guideline used. The physiological measures used

in this study were: Forced Vital Capacity (FVC),

Forced Expiratory Volume at the ﬁrst minute (FEV1),

Pick Expiratory Flow (PEF), Forced Expiratory Flow

(FEF), Maximum Voluntary Ventilation (VVM), Min-

imum Inspiratory Pressure (MIP), and Maximal Ex-

piratory Pressure (MEP). Besides those measures, the

variables gender, age, and body mass index were also

taken into account.

3.2 Identifying Inputs and Output

Variables

The physiological measures used were taken from

222 patients, being 80 with COPD (declared by phys-

iotherapists that work in rehabilitation program with

them based on clinical exams and spirometry). The

clinical and spirometry data derived two datasets for

training and testing of different ANN models con-

sidered in our experiments. The ﬁrst dataset (S1)

was created with the following variables as input:

age, gender, body mass index, FVC(o), FEV1(o),

PEF(o), FEF 25-75 %(o), Maximal Ventilatory Vol-

ume (MVV(o)), Maximal Inspiratory Pressure (MIP

(o)), and Maximal Expiratory Pressure (MEP(o)).

The notation (o) means the obtained values by the

spirometry. The second dataset (S2) was deﬁned by

the quotient between the measured and the expected

value for a normal subject (considering its gender and

age) following normality equations proposed by ATS,

being each input pattern composed of ﬁve variables:

FVC (%), FEV1(%), PEF(%), FEF25-75%(%), and

MVV(%). The input variables were normalized to

values between 0 and 1, including the variables age,

gender (binary digits 0 and 1) and body mass index.

As all ANN models used in this research are based

on supervised learning, the output answer of each in-

put is also included in the pattern representation. Two

classes were deﬁned: Normal subject and COPD sub-

ject, represented by the binary digits 0 and 1, respec-

tively.

3.3 ANN Architectures Analyzed

To devise this research three different ANN architec-

tures were designed to perform the classiﬁcation task,

namely: Linear Neuron with the Least Mean Square

(LMS) learning algorithm, Multilayer Perceptron and

Radial Basis Function ANN. This section describes

the modeling of each architecture and the results de-

scribing the performance of the ANNs are presented

in the next section.

Methodology for the Learning Process. For the

ANNs performance evaluation, a ten-fold cross-

validation method was used. Data was divided into

ten folds and the training was repeated ten times.

Each time, we applied nine folds for ANN training

and the remaining fold for validation. The ﬁnal accu-

racy was obtained by the average results over the ten

validation folds.

For all ANNs, the initial values of the synap-

tic weights were set to random values between

[-0.05,0.05].

To obtain a proper number of epochs for the ANNs

training, the following methodology was used: ini-

tially it was established an arbitrary upper bound for

the number of epochs. Then, for each epoch, it was

computed the mean square error (MSE) of the out-

puts, for the training set as well as for the validation

one. Analyzing the MSE curve of validation and the

MSE curve of training it was possible to determine

if the number of epochs was suitable or should in-

crease. Nevertheless, if during the training an indica-

tive of overtraining is detected (by analyzing the MSE

curve of validation), the training is promptly stopped.

This approach pointed out that an upper bound for the

number of training epochs equal to 100 was sufﬁcient

for training the three studied ANNs, considering our

datasets.

3.3.1 Linear ANN

This ANN is composed of only one linear neuron,

which output is given by the result of the linear com-

bination of synaptic weights and inputs, plus bias.

Here, this linear ANN was trained via the Least Mean

DiagnosingChronicObstructivePulmonaryDiseasewithArtificialNeuralNetworksusingHealthExpertGuidelines

209

Squares (LMS) algorithm developed by Widrow and

Hoff (Widrow and Hoff, 1960).

3.3.2 Multilayer Perceptron

The Multilayer Perceptron (MLP) is one of the most

studied ANN models. Its importance comes from the

addition of hidden layers and the use of a nonlinear

activation function (typically a sigmoid function) that

allows the ANN deal with problems which classes are

non-linearly separated. Both MLPs have one input

layer, one hidden layer and one output layer. The

number of neurons used at the input layer is the same

of the number of variables (10 to the S1 and 5 to the

S2) plus one for the bias. MLPs with different num-

bers of neurons in the hidden layer were tested, vary-

ing from 2 to 20, and that one with better performance

was selected. Evaluation criteria for the classiﬁers’

performance is presented in Section 3.4.

Training and Testing the MLPs. During the train-

ing phase, weights and bias of each MLP are updated

according to the scaled conjugate gradient (SCG)

method (Møller, 1993). SCG belongs to the class of

Conjugate Gradient Methods (CGMs) and has been

shown to be considerably faster than standard back-

propagation and other CGMs (Møller, 1993). Another

important advantage is that SCG is fully automated

including no user dependent parameters. The activa-

tion function used was the sigmoid logistic.

3.3.3 Radial Basis Function

A Radial Basis Function Neural Network (RBFNN)

is composed of an input layer, a hidden layer, and an

output layer. Just as in the two previous ANN models,

at the input layer, the number of neurons is equal to

the number of features. The hidden layer consists of

an arbitrary number of RBFs (e.g. Gaussian RBFs),

being each one deﬁned by a center position and a dis-

persion parameter (σ). The output layer is formed by

neurons that promote a linear combination of the ac-

tivations of the hidden layer neurons.

As for the previous ANNs, we designed two Ra-

dial Basis Function (RBF) networks for the classiﬁ-

cation of the sets S1 and S2, with input layer with 10

and 5 neurons, respectively, plus bias.

Training and testing the RBFNN. The RBFNN

training process was that one implemented by the

MATLAB Neural Network toolbox, which consists

of: (i) starting with no one neuron, the network is sim-

ulated; (ii) the input vector with the greatest error is

selected; (iii) a neuron is added to the hidden layer

with weights equal to that input vector and; ﬁnally,

(iv) the output layer weights are redesigned to min-

imize error. So, an upper bound value for the num-

ber of hidden layer neurons must be provided. Look-

ing for a better performance, a grid search procedure

was performed, varying the number of hidden neurons

from 5 to 30.

The values of the dispersion parameter σ were set

in an arbitrary way, receiving changes and combina-

tions at each simulation.

3.4 Evaluating Classiﬁers’ Performance

Four performance metrics were taken into considera-

tion in our analysis, namely: area under curve (AUC),

speciﬁcity, sensitivity and, accuracy.

AUC is obtained by the integration of the Receiver

Operating Characteristics (ROC) curve over a set of

thresholds considered. ROC is a technique to visual-

ize, organize and analyze the classiﬁer performance

by means of the False positive rate and True positive

rate relation, given a set of thresholds. Since AUC is a

portion of the area of the unit square, its value will al-

ways be between 0 and 1.0. AUC also has important

statistical properties and is related to other metrics,

such as Gini coefﬁcient (Fawcett, 2006).

Sensitivity and speciﬁcity are one approach to

quantifying the separation ability of the classiﬁer.

While the former is the proportion of true negatives

that are correctly identiﬁed by the test, the latter de-

scribes the proportion of true negatives that are cor-

rectly identiﬁed by the classiﬁer.

Accuracy, in turn, is deﬁned by the ratio of the

number of correct classiﬁcations to the number of pat-

terns tested.

Simulation results for MLP and RBFNN were

obtained using MATLAB R2010b Neural Network

Toolbox, whereas the linear ANN with LMS algo-

rithm was implemented by the authors on the same

platform.

4 RESULTS AND DISCUSSION

This section presents the results for simulations of the

three ANNs modeled. The results are presented in

terms of mean and standard deviation over 30 inde-

pendent runs. These results were obtained, consider-

ing a lower bound of acceptance of 0.5, i.e., if a lower

bound of 0.5 is established, outputs from 0.5 on are

classiﬁed as “1”.

HEALTHINF2013-InternationalConferenceonHealthInformatics

210

4.1 Results of Linear ANN

Tables 1 and 2 present the results for dataset S1 and

S2, respectively. The best results are those with high-

lighted background.

It was observed, from Table 1, that values of learn-

ing rate µ outside [0.01;0.7] decrease the performance

of the ANN, but inside this interval, the performance

keeps stable, achieving the best AUC equal to 0.971

(± 0.04), with µ = 0.5.

Table 1: Means and standard deviations of the four metrics

evaluated from the results of Linear ANN with dataset S1,

using different values for parameter µ.

µ AUC Accuracy Spec. Sens.

0.01

0.968 94.71% 89.29% 97.78%

(±0.04) (±4.45%) (±11.27%) (±4.43)

0.1

0.971 95.50% 91.29% 97.87%

(±0.04) (±4.03%) (±10.26%) (±3.53%)

0.3

0.965 95.79% 92.08% 97.88%

(±0.04) (±4.08%) (±9.85%) (±3.58%)

0.971 96.03% 92.88% 97.82%

0.5

(±0.04) (±3.69%) (±8.84%) (±3.71%)

0.7

0.957 93.64% 88.92% 96.29%

(±0.05) (±5.36%) (±12.65%) (±5.57%)

0.9

0.826 73.04% 77.08% 70.76%

(±0.16) (±18.15%) (±26.75%) (±32.56%)

Although the results of linear ANN for the dataset

S2 were stable for all values for the parameter µ an-

alyzed (Table 2), it is possible to observe a high fre-

quency of decrease in terms of accuracy and sensitiv-

ity. Conversely, we can observe an increase in terms

of speciﬁcity.

Table 2: Means and standard deviations of the four metrics

evaluated from the results of Linear ANN with dataset S2,

using different values for parameter µ.

µ AUC Accuracy Spec. Sens.

0.01

0.970 94.51% 96.00% 91.88%

(±0.03) (±4.41%) (±5.09%) (±9.97%)

0.1

0.965 94.19% 95.21% 92.38%

(±0.04) (±4.52%) (±5.43%) (±10.21%)

0.970 94.85% 95.73% 93.29%

0.3

(±0.04) (±4.45%) (±5.60%) (±8.57%)

0.5

0.968 94.13% 94.62% 93.25%

(±0.03) (±4.55%) (±5.85%) (±8.75%)

0.7

0.969 94.57% 95.01% 93.79%

(±0.04) (±4.48%) (±5.43%) (±8.64%)

0.9

0.968 94.41% 94.69% 93.92%

(±0.03) (±4.40%) (±5.85%) (±8.14%)

According to these good results achieved by the

linear model, it is possible to infer that with both sets

of selected variables, S1 and S2, lead to a nearly lin-

early separable classiﬁcation problem.

4.1.1 Results of MLP

Tables 3 and 4 present the results obtained by the Mul-

tilayer Perceptron with different number of hidden

neurons using datasets S1 and S2, respectively. The

best results are those with highlighted background.

As mentioned in Section 3.3, MLPs with different

numbers of neurons in the hidden layer were tested,

varying from 2 to 20. Nevertheless, we present the

results for those that led to the best performances.

Table 3: Means and standard deviations of the four metrics

evaluated from the results of MLP with dataset S1, consid-

ering different numbers of hidden neurons.

hidden

AUC Accuracy Spec. Sens.

neurons

0.963 95.21% 91.42% 97.34%

(±0.04) (±4.12%) (±9.04%) (±4.37%)

0.965 95.28% 91.29% 97.53%

(±0.04) (±3.91%) (±9.90%) (±3.97%)

0.963 95.16% 91.00% 97.51%

(±0.04) (±4.44%) (±10.73%) (±4.00%)

0.960 94.86% 90.12% 97.51%

(±0.05) (±4.51%) (±11.38%) (±3.92%)

0.966 95.63% 92.17% 97.58%

(±0.04) (±3.99%) (±9.21%) (±4.27%)

We can see that MLP also has produced good and

stable results for both datasets. While a more com-

plex model, with 9 neurons, presented the best perfor-

mance for dataset S1, a simpler one, with 3 neurons,

achieved the best results for dataset S2.

Table 4: Means and standard deviations of the four metrics

evaluated from the results of MLP with dataset S2, consid-

ering different numbers of hidden neurons.

hidden

AUC Accuracy Spec. Sens.

neurons

0.965 95.14% 95.97% 93.67%

(±0.04) (±4.33%) (±5.22%) (±8.82%)

0.962 94.57% 95.22% 93.42%

(±0.04) (±4.51%) (±5.62%) (±8.58%)

0.962 94.38% 95.32% 92.71%

(±0.04) (±4.82%) (±5.57%) (±10.25%)

0.956 93.85% 94.69% 92.33%

(±0.04) (±4.69%) (±6.10%) (±10.15%)

0.963 94.92% 95.51% 93.88%

(±0.04) (±4.18%) (±5.47%) (±9.06%)

Even presenting similar performance in terms of

AUC for both dataset S1 and S2, we can identify de-

creases in the accuracy and sensitivity in dataset S2,

when compared with results obtained with dataset S1.

On the other hand, it is possible to notice an increment

in terms of speciﬁcity.

DiagnosingChronicObstructivePulmonaryDiseasewithArtificialNeuralNetworksusingHealthExpertGuidelines

211

4.1.2 Results of RBFNN

Tables 5 and 6 present the results of some combina-

tions of values to the parameters σ, using RBFNN

with dataset S1 and S2, respectively. As informed in

Section 3.3, we performed a grid search procedure to

ﬁnd out a proper number of hidden neurons, varying

from 5 to 30. For each conﬁguration, we varied the

dispersion parameter σ from 0.1 to 10. An RBFNN

with 15 hidden neurons has shown the best results.

Due to lack of space, we omitted the results of the

other experiments.

Table 5: Means and standard deviations of the four metrics

evaluated from the results of RBFNN with dataset S1, using

different values for parameter σ.

σ AUC Accuracy Spec. Sens.

0.1

0.908 89.05% 79.71% 94.32%

(±0.06) (±6.04%) (±16.04%) (±6.49%)

0.5

0.966 95.95% 92.58% 97.84%

(±0.04) (±3.77%) (±8.87%) (±3.69%)

1.0

0.969 96.21% 92.25% 98.44%

(±0.04) (±4.02%) (±9.61%) (±3.10%)

0.971 95.93% 92.42% 97.91%

(±0.03) (±4.00%) (±9.41%) (±3.71%)

Regarding the results presented in Table 5, higher

values of the dispersion parameter, particularly 1 and

10, presented the best performances in terms of AUC

and sensitivity.

Table 6: Means and standard deviations of the four metrics

evaluated from the results of RBFNN with dataset S2, using

different values for parameter σ.

σ AUC Accuracy Spec. Sens.

0.968 96.24% 96.11% 96.46%

0.1

(±0.04) (±3.63%) (±4.95%) (±6.74%)

0.5

0.964 95.61% 95.21% 96.33%

(±0.04) (±4.07%) (±5.75%) (±6.63%)

1.0

0.960 95.16% 95.03% 95.38%

(±0.05) (±4.27%) (±5.78%) (±7.30%)

0.961 95.09% 95.05% 95.17%

(±0.05) (±4.39%) (±5.63%) (±7.62%)

As already observed in the two previous ANN

models, the RBFNN also achieved similar perfor-

mance in terms of AUC for both dataset S1 and S2,

and a slight decrease in accuracy and sensitivity in

dataset S2, when compared with results obtained with

dataset S1. (Exception for σ = 0.1). Conversely, we

notice an increment in terms of speciﬁcity.

4.2 Comparing Models

In this section we present a statistical comparative

study of the three ANN architectures analyzed. This

study is conducted only looking at the values of AUC

and sensitivity. The reason for this is that AUC carries

more information about the classiﬁer power, whereas

sensitivity is related to false negative rate - fnr (type

II error), through sens = (1 − f nr), which quantiﬁes

cases where a patient is diagnosed as normal when

he actually is not, being so, a dangerous diagnosis.

Therefore, classiﬁers with high sensitivity (low fnr)

are preferable.

We carried out a one-tailed Student t -test with

a signiﬁcance of 0.05 to compare the performance

of the classiﬁers which found the best results (high-

lighted background in tables) and the results are pre-

sented in Table 7.

Table 7: The t-test results regarding Alg.1 – Alg.2 is shown

as “+”, “–”, or “∼” when Alg. 1 is signiﬁcantly better than,

signiﬁcantly worse than, or statistically equivalent to Alg.

2, respectively.

t-test result

S1 S2

AUC Sens. AUC Sens.

LMS – RBF ∼ – ∼ –

RBF – MLP ∼ + ∼ +

MLP – LMS ∼ ∼ ∼ ∼

No signiﬁcant difference between linear ANN

(LMS) and MLP was found for both AUC and sen-

sitivity. It is worth mentioning that MLP is a more

complex model and require a more computational in-

tensive learning process than the least mean square

algorithm.

RBFNN also did not present signiﬁcant better per-

formance when compared to MLP and linear ANN in

terms of AUC, but produced classiﬁers with the high-

est sensitivity. Given these results, we could say that

RBFNN seems to be the ﬁrst option to apply to the

COPD identiﬁcation problem.

4.3 Comparing Feature Sets

We performed a statistical comparative study of

the classiﬁers’ performance over the two different

datasets, S1 and S2. The former, as described in

Section 3.2, consists of normalized raw data of three

non-physiological and seven physiological measures.

The latter considers ﬁve physiological measures that

incorporate previous knowledge from health experts.

The results are presented in Table 8.

It is possible to see that all classiﬁers lose the ca-

pacity of reduce false negatives when using the set of

HEALTHINF2013-InternationalConferenceonHealthInformatics

212

Table 8: The t-test results regarding Alg. is shown as “+”

or “∼” when the performance of the Alg. with S1 is signif-

icantly better than or statistically equivalent to the perfor-

mance when using S2, respectively.

AUC Sensitivity

LMS ∼ +

MLP ∼ +

RBF ∼ +

variables S2. One hypothesis is that using predicted

values from regression equations as reference values,

may add noise to the physiological measure, due to

residuals from the regression, which, in turn, can mis-

lead the classiﬁer.

4.4 Identifying Discriminating

Variables

We are now interested in ﬁnding out what are the most

discriminating variables. This information could pro-

vide some clues about what physical and physiologic

characteristics have more inﬂuence in the decision

process implemented by the ANNs studied.

As the three analyzed classiﬁers achieved quite

similar performance, we will look at the weights of

the linear ANN learned through the least mean square

algorithm. The means and standard deviations of the

weights obtained by linear ANN for dataset S1 and

S2, based on 30 independent runs, are depicted in Fig-

ures 1 and 2, respectively.

−3

−2

−1

weights

Age

Gender

BMI

FVC(o)

FEV1(o)

PEF(o)

FEF25−75(o)

MVV(o)

MIP(o)

MEP(o)

Figure 1: Means and standard deviations of the linear ANN

weights for the S1 dataset over 30 executions.

For the dataset S1, it is possible to see that gen-

der and body mass index of a subject have a little

inﬂuence in the decision process. Additionally, the

physiologic measure FEF25-75 (o) are practically ig-

nored by the classiﬁer, although it is worth mention-

ing its high standard deviation. On the other hand, the

most discriminating variables refer to the physiologi-

cal measures FVC (o), FEV1(o), and PEF(o).

−4.5

−4

−3.5

−3

−2.5

−2

−1.5

−1

−0.5

0.5

weights

FVC(%)

FEV1(%)

PEF(o)

FEF25−75(%)

MVV(%)

Figure 2: Means and standard deviations of the linear ANN

weights for the S2 dataset over 30 executions.

The only neglected variable by the linear ANN in

the S2 dataset is the physiologic measure FVC(%),

but also has a high standard deviation. That using ref-

erence measures, sometimes obtained through regres-

sion equation, to guide the physiological measures led

to a different classiﬁcation problem, probably more

difﬁcult to deal with, given the results.

5 CONCLUDING REMARKS AND

FUTURE WORK

The analysis and classiﬁcation of physiological mea-

sures is not a trivial task. The complexity to develop

this kind of research includes accurate data acquisi-

tion, selection of a suitable classiﬁer that can provide

a good representation of that patterns. This work pre-

sented results of modeling, training and testing three

different ANN architectures to identify the Chronic

Obstructive Pulmonary Disease. Our results highlight

the potential of ANNs as a support decision tool for

the problem of COPD identiﬁcation.

The results showed that even simple models as the

linear neuron with LMS algorithm had a good perfor-

mance. It allows us to infer that these sets of selected

variables, S1 and S2, conduct the COPD identiﬁca-

tion problem to a nearly linearly separable classiﬁca-

tion problem.

Two sets of variables were considered in our anal-

ysis: one composed of normalized raw data of three

non-physiological (age, gender and body mass index)

and seven spirometric measures; and the other con-

sisting of ﬁve variables deﬁned as the ratio of the

DiagnosingChronicObstructivePulmonaryDiseasewithArtificialNeuralNetworksusingHealthExpertGuidelines

213

spirometric measures to the expected value for a nor-

mal subject following the guideline provided by the

American Thoracic Society. It was performed as an

attempt to incorporate knowledge from health experts

to the dataset. Our results showed that, despite hav-

ing similar performance in terms of AUC, all classi-

ﬁers lose the capacity of reduce false negatives when

using the second set of variables.

Among the ANNs models analyzed, RBFNNs ob-

tained similar results in terms of classiﬁcation power,

but better performance when looking at the classi-

ﬁers’ sensitivity, for both datasets. This measure tells

us that RBFNN classiﬁers are more likely to avoid

false negative diagnosis, i.e., cases when a COPD pa-

tient is diagnosed as normal, that may be dangerous.

Such results agree with that obtained by (Mehrabi

et al., 2009) and the performance measures obtained

in our work are slight better, even considering that

other feature set had been considered. The results

obtained with the application of ANN in the classi-

ﬁcation of diseases encourage the study of new ap-

plications of such models to help with problems of

biomedicine, pointing out the ANN as a powerful

technique to help with the understanding and diag-

nosing diseases.

In this work, ANNs were used only to identify the

presence or absence of COPD. As future work, it is

intended to apply ANN to classify the level of sever-

ity of the disease as well as to support decision on

treatment, according to this level.

ACKNOWLEDGEMENTS

This work was supported in part by the National

Council for Research and Development of Brazil. We

also thank Prof. Dr. Antonio Fernando Brunetto

(in Memoriam) by his collaboration as the expert in

COPD rehabilitation.

REFERENCES

Azeredo, C. A. C. (2002). Fisioterapia Respirat

oria Mod-

erna. Manole, S

ao Paulo, 4 edition.

Er, O. and Temurtas, F. (2008). A study on chronic ob-

structive pulmonary disease diagnosis using multi-

layer neural networks. Journal of Medical Systems,

32:429–432.

Er, O., Yumusak, N., and Temurtas, F. (2010). Chest dis-

eases diagnosis using artiﬁcial neural networks. Ex-

pert Systems with Applications, 37(12):7648–7655.

Fawcett, T. (2006). An introduction to ROC analysis. Pat-

tern Recogn. Lett., 27(8):861–874.

Fontenla-Romero, O., Guijarro-Berdinas, B., Alonso-

Betanzos, A., and Moret-Bonillo, V. (2005). A new

method for sleep apnea classiﬁcation using wavelets

and feedforward neural networks. Artiﬁcial Intelli-

gence in Medicine, 34:65–76.

Group, A. W. (2004). Update to the Latin American Tho-

racic Society (ALAT) Recommendations on Infec-

tious Exacerbation of COPD. Archivos de Bronconeu-

mologia, 40:315–325.

Lenfant, C. (1998). Estrat

egia global para o diagn

ostico, a

conduta e a prevenc¸

ao da doenc¸a pulmonar obstrutiva

onica. Technical report, GOLD - Global Initiative

for Chronic Obstructive Lung Disease.

Mehrabi, S., Maghsoudloo, M., Arabalibeik, H., Noor-

mand, R., and Nozari, Y. (2009). Application of mul-

tilayer perceptron and radial basis function neural net-

works in differentiating between chronic obstructive

pulmonary and congestive heart failure diseases. Ex-

pert Systems with Applications, 36(3, Part 2):6956–

6959.

Møller, M. F. (1993). A scaled conjugate gradient algo-

rithm for fast supervised learning. Neural Networks,

6(4):525–533.

Nathell, L., Nathell, M., Malmberg, P., and Larsson, K.

(2007). COPD diagnosis related to different guide-

lines and spirometry techniques. Respiratory Re-

search, 8.

Passold, F., Garcia Ojeda, R., and Muniz Barreto, J. (1996).

Hybrid expert system in anesthesiology for critical pa-

tients. In Electrotechnical Conference, 1996. MELE-

CON ’96., 8th Mediterranean, volume 3, pages 1486–

1489.

Wadie, B. S., Badawi, A. M., Abdelwahed, M., and El-

emabay, S. M. (2006). Application of artiﬁcial neu-

ral network in prediction of bladder outlet obstruction:

A model based on objective, noninvasive parameters.

Urology, 68:1211–1214.

Widrow, B. and Hoff, M. E. (1960). Adaptive switching

circuits. In IRE WESCON Convention Record, Part 4,

pages 96–104, New York. IRE.

Yan, H., Jiang, Y., Zheng, J., Peng, C., and Li, Q. (2006). A

multilayer perceptron-based medical decision support

system for heart disease diagnosis. Expert Systems

with Applications, 30(2):272–281.

Zhou, Z.-H. and Jiang, Y. (2003). Medical diagnosis with

c4.5 rule preceded by artiﬁcial neural network ensem-

ble. IEEE Transactions on Information Technology in

Biomedicine, 2003, 7(1): 37-42, 7(1):37–42.

HEALTHINF2013-InternationalConferenceonHealthInformatics

214