Вы находитесь на странице: 1из 9

Fruit Fly Optimization Algorithm Based SVM Classifier

for Efficient Detection of Parkinsons Disease

Liming Shen, Huiling Chen(), Wenchang Kang,


Haoyue Gu, Bingyu Zhang, and Ting Ge

College of Physics and Electronic Information, Wenzhou University,


Wenzhou 325035, Peoples Republic of China
chenhuiling_jsj@wzu.edu.cn

Abstract. In this paper, we present a fruit fly optimization algorithm (FOA)


based support vector machine (SVM) classification scheme, termed as FOA-
SVM, and it is applied successfully to Parkinsons disease (PD) diagnosis. In
the proposed FOA-SVM, the set of parameters in SVM is tackled efficiently by
the FOA technique. The effectiveness and efficiency of FOA-SVM has been ri-
gorously evaluated against the PD dataset by comparing with the particle swarm
optimization algorithm (PSO) optimized SVM (PSO-SVM), and grid search
technique based SVM (Grid-SVM). The experimental results demonstrate that
the proposed approach outperforms the other two counterparts in terms of diag-
nosis accuracy as well as the fewer CPU time. Promisingly, the proposed me-
thod can be regarded as a useful clinical decision tool for the physicians.

Keywords: Support vector machine Parameter optimization Fruit fly optimi-


zation Parkinsons disease diagnosis Medical diagnosis

1 Introduction

Parkinsons disease (PD) is one kind of degenerative diseases of the nervous system,
it has become the second most common degenerative disorders of the central nervous
system after Alzheimer's disease [1]. Till now, the cause of PD hasnt been unco-
vered. However, it is possible to alleviate symptoms significantly at the onset of the
illness in the early stage [2]. It has also been proven that a vocal disorder may be one
of the first symptoms to appear nearly 5 year before clinical diagnose [3]. The vocal
impairment symptoms related with PD are known as dysphonia (inability to produce
normal vocal sounds) and dysarthria (difficulty in pronouncing words) [4]. Therefore,
dysphonic indicators may play essential role in the early stage of PD diagnosis. Little
et al [5] have made the first attempt to utilize the dysphonic indicators in their study
to help discriminate PD patients from healthy ones. In their study, support vector
machine (SVM) in combination with the feature selection approach was taken to di-
agnose PD, the simulation results has shown that the proposed method can discrimi-
nate PD patients from healthy ones with approximately 90% classification accuracy
using only four dysphonic features. After then, various techniques have been devel-

Springer International Publishing Switzerland 2015


Y. Tan et al. (Eds.): ICSI-CCI 2015, Part II, LNCS 9141, pp. 98106, 2015.
DOI: 10.1007/978-3-319-20472-7_11
Fruit Fly Optimization Algorithm Based SVM Classifier for Efficient Detection 99

oped to study the PD diagnosis problem from the perspective of dysphonic indicators,
including Artificial Neural Networks (ANNs) [6, 7], SVM [8, 9], Dirichlet process
mixtures [10], multi-kernel relevance vector machines [11], similarity classifier [12],
rotation forest [13], fuzzy k-nearest neighbor (FKNN) [14].
Among the proposed methods, SVM has shown to be a very promising tool for di-
agnosing PD. However, in our opinion despite its great potential, SVM has not re-
ceived the attention it deserves in the PD diagnosis literature as compared to other
research fields. SVM was first introduced by Vapnik [15], which has many good
properties and has found its application in many fields. However, it has been pointed
out that model parameter setting has great impact on the performance of SVM [16].
Values of parameters such as penalty parameter C and the kernel parameter g of the
kernel function should be properly tuned before SVM applying to the practical prob-
lems. Traditionally, these parameters were handled by the grid-search method and the
gradient descent method. However, one common drawback of these methods is that
they are vulnerable to local optimum. Recently, biologically inspired global optimiza-
tion methods such as genetic algorithm and particle swarm optimization (PSO) have
been considered to have a better chance of finding the global optimum solution than
the traditional aforementioned methods. As a new member of the swarm-intelligence
algorithms, fruit fly optimization algorithm (FOA) [17] has been found to be a useful
tool for real-world optimization problems such as the semiconductor final testing
scheduling problem [18], continuous function optimization problems [19], parameter
optimization of generalized regression neural network [20] and Least Squares SVM
[21] for regression problems. This study attempts to employ FOA to tackle the para-
meter optimization of SVM and applied the resultant model for effective detection of
PD. To the best of our knowledge, FOA has not been utilized to optimize the parame-
ters of SVM classifier. Therefore, this study will be the first to report the FOA opti-
mized SVM classifier and its application to PD diagnosis.
The remainder of this paper is organized as follows. The detailed implementation
of the FOA-SVM method is presented in section 2. Section 3 describes the experi-
mental design. The experimental results and discussion of the proposed approach are
presented in Section 4. Finally, Conclusions and recommendations for future work are
summarized in Section 5.

2 Proposed FOA-SVM Model

This study proposes a novel FOA-SVM model for parameter optimization problem of
SVM. The proposed model was comprised of two procedures as shown in Figure 1,
the one is the inner parameter optimization, and the other is the outer performance
evaluation. During the inner parameter optimization procedure, the parameters of
SVM are adjusted dynamically by the FOA technique via the 5-fold cross validation
100 L. Shen et al.

Fig. 1. The flowchart of the proposed FOA-SVM diagnostic system

(CV) strategy. And then the obtained optimal parameters are fed to SVM prediction
model to perform the classification task for Parkinsons disease diagnosis in the outer
loop using the 10-fold CV strategy.
Fruit Fly Optimization Algorithm Based SVM Classifier for Efficient Detection 101

The classification accuracy is taken into account in designing the fitness:

f = avgACC = (iK=1testACCi ) k (1)

where avgACC in the function f represents the average test accuracy achieved by the
SVM classifier via 5-fold CV strategy.

3 Experimental Studies

3.1 Data Description

The Parkinsons data was taken from UCI machine learning repository
(http://archive.ics.uci.edu/ml/datasets/Parkinsons, last accessed: December 2014). The
objective of this dataset is to discriminate healthy people from those with Parkinson's
disease (PD). In the medical experiment, various biomedical voice measurements
were recorded for 23 patients with PD and 8 healthy controls. The time since diagnos-
es ranged from 0 to 28 years, and the ages of the subjects ranged from 46 to 85 years,
with a mean age of 65.8. Each subject provides an average of six phonations of the
vowel (yielding 195 samples in total), each 36 seconds in length. It should be noted
that there is no missing values in the dataset, and the whole features are real valued.

3.2 Experimental Setup

The proposed FOA-SVM classification model was implemented using MATLAB plat-
form. For SVM, LIBSVM implementation was utilized, which was originally developed
by Chang and Lin [22]. We implemented the FOA algorithm from scratch. The computa-
tional analysis was conducted on Windows 7 operating system with AMD Athlon 64 X2
Dual Core Processor 5000+ (2.6 GHz) and 4GB of RAM. Before constructing the SVM
models, the data was scaled to the range of [0, 1] to avoid the feature values in greater
numerical ranges dominating those in smaller numerical ranges.

3.3 Measure for Performance Evaluation

Classification accuracy (ACC), area under the receiver operating characteristic curve
(AUC) criterion, sensitivity and specificity were used to test the performance of the
proposed FOA-SVM model. ACC, sensitivity and specificity are defined as follows:

Accuracy = TP + TN / (TP + FP + FN + TN ) 100% (2)


Sensitivity = TP / (TP + FN ) 100% (3)
Specificity = TN / ( FP + TN ) 100% (4)
where TP is the number of true positives, FN is the number of false negatives, TN is
the number of true negatives, and FP is the number of false positives. AUC is the area
under the ROC curve, which is one of the best methods for comparing classifiers in
102 L. Shen et al.

two-class problems. in this study the method proposed in [23] was implemented to
compute the AUC.

4 Experimental Results and Discussions

The swarm size and number of generations play important role in controlling the
search ability of FOA. Thus, we firstly investigated the impact of the three factors on
the performance of FOA. Different sizes of swarm from 10 to 22 with step size of 2
were evaluated when the generation is fixed to 60, the detailed results is presented in
Table 1. From the table, we can see that the best performance was achieved when the
swarm size is 20, where the ACC, AUC, sensitivity and specificity are 95.95%,
92.81%, 98.52% and 87.10%, respectively. When the size of 20 is fixed, different
generations from 20 to 120 with step size of 20 were tried. As shown in Table 2, we
can see that the best performance of FOA-SVM is achieved when the generations are

Table 1. The detailed results of FOA-SVM with different swarm size

Swarm size FOA-SVM


sizepop (when
ACC (%) AUC (%) Sensitivity (%) Specificity (%)
gen = 60)
10 94.34(5.08) 91.06(7.99) 97.95(3.30) 84.17(15.02)
12 93.87(4.06) 88.58(7.64) 98.66(2.83) 78.50(15.58)
14 94.79(6.57) 91.95(7.62) 99.33(2.11) 84.57(14.62)
16 94.87(4.23) 91.54(8.57) 97.94(3.34) 85.14(17.94)
18 95.37(5.17) 92.18(11.59) 98.71(2.72) 85.65(23.17)
20 95.95(4.61) 92.81(9.32) 98.52(3.13) 87.10(19.50)
22 95.34(5.73) 91.54(122) 99.44(1.76) 83.86(19.32)
The best results have been shown in bold.

Table 2. The detailed results of FOA-SVM with different generations

Generation FOA-SVM
Gen (when Sensitivity Specificity
sizepop = 20) ACC (%) AUC (%)
(%) (%)
20 93.32(4.19) 85.92(8.86) 99.23(2.43) 72.62(17.5)
40 94.39(5.98) 91.75(8.47) 97.91(3.41) 85.58(16.52)
60 95.84(5.92) 94.75(7.02) 99.09(2.87) 90.42(13.54)
80 94.39(3.68) 92.05(9.81) 98.27(1.86) 85.83(19.18)
100 93.87(5.31) 92.01(6.18) 97.62(5.50) 86.39(13.95)
120 93.84(4.81) 89.04(7.16) 98.75(3.95) 79.33(12.43)
The best results have been shown in bold.
Fruit Fly Optimization Algorithm Based SVM Classifier for Efficient Detection 103

set to 60 with the ACC of 95.84%, AUC of 94.75%, sensitivity of 99.09% and speci-
ficity of 90.42%. In the above two tables, the average results of 10-fold CV are pre-
sented with the standard deviation described in the parenthesis. From the above analy-
sis, we can see that FOA-SVM reaches the best performance when swarm size=20
and generations=60 in terms of ACC, AUC, sensitivity and specificity. Therefore,
these parameter values are adopted for the proposed FOA-SVM to implement the
subsequent experiments.
Apart from the FOA-SVM classifier, PSO-SVM and Grid-SVM classifiers were
implemented for the comparison purpose. For PSO-SVM, the number of the iterations
and particles are set the same as that of FOA-SVM, they are 60 and 20, respectively.
The maximum velocity vmax is set about 60% of the dynamic range of the variable on
each dimension for the continuous type of dimensions, acceleration coefficients c1 =
2, c2 = 2, maximum and minimum values of the inertia weight wmax and wmin are set to
0.9 and 0.4, respectively. The searching ranges of C [2 ^ (5), 2 ^ (15)] and g
[2 ^ (15), 2 ^ (3)] for PSO-SVM and Grid-SVM were set as the same. Table 3
shows the detailed results of ACC, AUC, sensitivity, specificity, and optimal pairs of
(C, g) for each fold obtained by FOA-SVM. The detailed comparison results among
PSO-SVM, Grid-SVM and FOA-SVM in terms of ACC, AUC, sensitivity and speci-
ficity are shown in Figure 2. It can be observed from Figure 2 that the performance of
FOA-SVM is superior over the other two competitors in most folds.

Table 3. The detailed results of FOA-SVM on the PD dataset

Fold FOA-SVM
No. Accuracy AUC Sensitivity Specificity C g
1 1.0000 1.0000 1.0000 1.0000 1.09637 1.09637
2 1.0000 1.0000 1.0000 1.0000 1.27755 1.27755
3 0.9474 0.9000 1.0000 0.8000 1.7315 1.7315
4 1.0000 1.0000 1.0000 1.0000 1.23075 1.23075
5 1.0000 1.0000 1.0000 1.0000 1.13659 0.988679
6 1.0000 1.0000 1.0000 1.0000 1.07801 1.07801
7 0.8421 0.8500 1.0000 0.7000 1.6111 1.6111
8 1.0000 1.0000 1.0000 1.0000 1.30109 1.30109
9 0.9000 0.8333 1.0000 0.6667 2.34289 2.34289
10 0.8947 0.8920 0.9091 0.8750 2.18101 2.18101
Avg. 0.9584 0.9475 0.9909 0.9042 1.49869 1.48390
Dev. 0.0592 0.0702 0.0287 0.1354 0.45641 0.47159
104 L. Shen et al.

Fig. 2. ACC, AUC, sensitivity and specificity obtained for each fold by PSO-SVM, Grid-SVM
and FOA-SVM

Fig. 3. The comparison results of CPU time for the three methods

In order to investigate the efficiency of the proposed method, we have compared


FOA-SVM with other two methods in terms of CPU time. As shown in Figure 3,
FOA-SVM needs almost the same CPU times as that of Grid-SVM, PSO-SVM ap-
pears to be the most time consuming among the three methods. Since the grid search
technique needs no iterative procedure, it is comparatively fast when the search
ranges are set properly. It is important to note that FOA-SVM needs only about 12.26
seconds on average during the whole 10 folds CV procedure, even when there are so
Fruit Fly Optimization Algorithm Based SVM Classifier for Efficient Detection 105

many generations involved. However, PSO-SVM consumed 91.79 seconds when the
same generations and swarm size were considered. From the above analysis, we can
see that FOA-SVM has the evident superiority over PSO-SVM in terms of CPU time
as well as the classification performance.

5 Conclusions and Future Work

This work has explored a new method, FOA-SVM, for effective and efficient detec-
tion of PD. The main novelty of this paper lies in the proposed FOA-based approach,
which aims at maximizing the generalization capability of the SVM classifier by ex-
ploring the new swarm intelligence technique for optimal parameter tuning for PD
diagnosis. The empirical experiments have demonstrated the evident superiority of the
proposed FOA-SVM over PSO-SVM and Grid-SVM. It indicates that the proposed
FOA-SVM method can be utilized as a valuable alternative clinical solution to PD
diagnosis. In the future work, we plan to apply the proposed method to other medical
diagnosis problems.

Acknowledgments. This research is supported by the National Natural Science Foundation


of China (NSFC) under Grant Nos. of 61303113 and 61402337. This work is also funded by
Zhejiang Provincial Natural Science Foundation of China under Grant Nos. of LQ13G010007,
LQ13F020011 and LY14F020035.

References

1. de Lau, L.M.L., Breteler, M.M.B.: Epidemiology of Parkinsons disease. The Lancet


Neurology 5(6), 525535 (2006)
2. Singh, N., Pillay, V., Choonara, Y.E.: Advances in the treatment of Parkinsons disease.
Progress in Neurobiology 81(1), 2944 (2007)
3. Harel, B., Cannizzaro, M., Snyder, P.J.: Variability in fundamental frequency during
speech in prodromal and incipient Parkinsons disease: A longitudinal case study. Brain
and Cognition 56(1), 2429 (2004)
4. Baken, R.J., Orlikoff, R.F.: Clinical measurement of speech and voice, 2nd edn. Singular
Publishing Group, San Diego (2000)
5. Little, M.A., et al.: Suitability of Dysphonia Measurements for Telemonitoring of Parkin-
sons Disease. IEEE Transactions on Biomedical Engineering 56(4), 10151022 (2009)
6. Das, R.: A comparison of multiple classification methods for diagnosis of Parkinson dis-
ease. Expert Systems with Applications 37(2), 15681572 (2010)
7. Astrm, F., Koker, R.: A parallel neural network approach to prediction of Parkinsons
Disease. Expert Systems with Applications 38(10), 1247012474 (2011)
8. Sakar, C.O., Kursun, O.: Telediagnosis of Parkinsons Disease Using Measurements of
Dysphonia. Journal of Medical Systems 34(4), 19 (2010)
9. Li, D.C., Liu, C.W., Hu, S.C.: A fuzzy-based data transformation for feature extraction to
increase classification performance with small medical data sets. Artificial Intelligence in
Medicine 52(1), 4552 (2011)
106 L. Shen et al.

10. Shahbaba, B., Neal, R.: Nonlinear models using Dirichlet process mixtures. The Journal of
Machine Learning Research 10, 18291850 (2009)
11. Psorakis, I., Damoulas, T., Girolami, M.A.: Multiclass Relevance Vector Machines: Spar-
sity and Accuracy. IEEE Transactions on Neural Networks 21(10), 15881598 (2010)
12. Luukka, P.: Feature selection using fuzzy entropy measures with similarity classifier.
Expert Systems with Applications 38(4), 46004607 (2011)
13. Ozcift, A., Gulten, A.: Classifier ensemble construction with rotation forest to improve
medical diagnosis performance of machine learning algorithms. Comput. Methods
Programs Biomed. 104(3), 443451 (2011)
14. Chen, H.-L., et al.: An efficient diagnosis system for detection of Parkinsons disease
using fuzzy k-nearest neighbor approach. Expert Systems with Applications 40(1),
263271 (2013)
15. Vapnik, V.N.: The nature of statistical learning theory. Springer, New York (1995)
16. Keerthi, S.S., Lin, C.-J.: Asymptotic behaviors of support vector machines with Gaussian
kernel. Neural Computation 15(7), 16671689 (2003)
17. Pan, W.-T.: A new Fruit Fly Optimization Algorithm: Taking the financial distress model
as an example. Knowledge-Based Systems 26, 6974 (2012)
18. Zheng, X.-L., Wang, L., Wang, S.-Y.: A novel fruit fly optimization algorithm for the
semiconductor final testing scheduling problem. Knowledge-Based Systems 57, 95103
(2014)
19. Pan, Q.-K., et al.: An improved fruit fly optimization algorithm for continuous function
optimization problems. Knowledge-Based Systems 62, 6983 (2014)
20. Li, H.-Z., et al.: A hybrid annual power load forecasting model based on generalized
regression neural network with fruit fly optimization algorithm. Knowledge-Based
Systems 37, 378387 (2013)
21. Li, H., et al.: Annual Electric Load Forecasting by a Least Squares Support Vector
Machine with a Fruit Fly Optimization Algorithm. Energies 5(11), 44304445 (2012)
22. Chang, C.-C., Lin, C.-J.: LIBSVM: A library for support vector machines. ACM Trans.
Intell. Syst. Technol. 2(3), 127 (2011)
23. Fawcett, T.: ROC graphs: Notes and practical considerations for researchers. Machine
Learning 31, 138 (2004)

Вам также может понравиться