Biomedical Research

Abstract

In this paper, a new hybrid classification approach, which uses Weighted-Particle Swarm Optimization (WPSO) for data clustering in sequence with Smooth Support Vector Machine (SSVM) for classification is proposed. The performance of WPSO clustering is compared with K means and fuzzy methods using intercluster, intracluster and validity index. The accuracy of proposed WPSO-SSVM classification methodology are 83.76% for liver disorder, 98.42% for WBCD, 95.21% for mammographic mass data which are better than in existing literature.

Keywords

Smooth support vector machine (SSVM), Particle swarm optimization (PSO), Clustering, Classification.

Introduction

Medical data mining has a great potential for exploring hidden patterns and extracting useful information for decision support [1]. Benefits of introducing machine learning into medical analysis are to increase diagnostic accuracy, reduce costs and human resources [2]. Case based reasoning [3] process is an approach for developing knowledge-based medical decision support system which solves new problems based on the solutions of similar past problems.

Materials and Methods

Assume a medical library with each case in the library as index of corresponding features (e₁, e₂, ..., e_N) having an associated action, with collection of features F_j (j=1…. n) representing the cases and variable V denoting the action. The i^th case e_j in the library can be represented as an n+1-dimensional vector, i.e. e_i=(x_i1, x_i2, ......, x_in, y_i). Where x_ij corresponds to the value of feature F_j (j=1... n) and y_i corresponds to the action (i=1... n). If for each j (1 ≤ j ≤ n) a weight w_j (w_j (0, 1)) has been assigned to the j^th feature to indicate the importance of the feature, then for any pair of cases e_p and e_q in the library, a weighted distance metric d_pq^(w) is defined as:

Where x_pj is the p^th case with j^th feature and xqj is the qth case with j^th feature. Using the weighted distance a similarity measure SM_pq^(w) is calculated usingSM_pq^(w)=1/(1+αd_pq^(w))

Where α is a positive parameter. The weighted feature assignment algorithm is presented in Figure 1.

Figure 1: Weighed feature assignment algorithm.

PSO is a population-based search algorithm and each particle is associated with a velocity and its algorithm is presented in Figure 2.

Figure 2: PSO based clustering algorithm.

A nonlinear version of the SSVM [4] is used for classification of datasets after clustering.

Results

The WBCD, mammographic mass and liver disorder dataset are obtained from UCI machine learning repository [5]. Weighed PSO clustering is applied on the datasets (Figure 3) and compared with K-means and FCM in terms of intercluster, intra cluster and validity index as shown in Table 1. The inter cluster distance of any two cluster should be high which is best for PSO as seen in Table 1. Intra cluster means the compactness of a cluster and its value should be least as possible and is again best for PSO.

Table 1: Comparison of inter, intra and validity index with FCM, K-means and PSO for breast cancer (WBCD) and Liver disorder dataset [6].

Figure 3: Weighed PSO based clustering.

The clustered output is classified using SSVM using fivefold cross validation in which randomly split database is averaged to provide the best indication of true classification performance and the performance comparison of datasets is presented in Table 2 and accuracy is shown in Figure 4.

Table 2: Performance comparison of datasets.

Figure 4: Accuracy (%) comparison of proposed method with other existing methods in literature.

Conclusions

This paper proposes a new WPSO-SSVM technique to improve the classification accuracy of medical datasets and the obtained results are found to outperform all the present stateof- art classifiers existing in literature. The future work will be to test the proposed technique in other benchmark datasets to prove the robustness of the proposed algorithm.

References

1. Pei CC, Jyun JL, Chen HL. An attribute weight assignment and particle swarm optimization algorithm for medical database classifications. Comp Met Prog Biomed 2011; 1-11.
2. Bojarczuk HSL, Freitas AA. Genetic programming for knowledge discovery in chest-pain diagnosis: exploring a promising data mining approach. IEEE Eng Med BiolMagaz 2000; 19: 38- 44.
3. Klaus DA. Case-based reasoning for medical decision support tasks: the inreca approach. Artificial Intel Med 1998; 12: 25-41.
4. Yuh-Jye L,Mangasarian. Smooth support vector machine. Comp OptimizAppl2001; 20: 5-22.
5. Data Mining Institute, University of Wisconsin, Technical Report 99-03.
6. https://archive.ics.uci.edu/ml/datasets.html (Liver disorders, Mammographic mass and Wisconsin Breast Cancer (Original)).