Application of Genetic Algorithm for Cancer Diagnosis by Feature Selection

Application of Genetic Algorithm for Cancer Diagnosis by Feature Selection

Tanzeem Khan Mansoori

1. Tech. scholar CSE LNCTS Bhopal, India

   Prof. Amrit Suman

   CSE Dept.

   LNCTS Bhopal, India

   Dr. Sadhna K. Mishra

   CSE Dept.

   LNCTS Bhopal, India

   Abstract The main aim of feature selection is to select a set of input variables by removing features with very little or no analytical information. Feature selection strategies are often decomposed into three broad categories. View of the substantial range of existing feature selection algorithms, the necessity arises to count on criteria which enable to adequately decide in certain situations which algorithm should be used. This work reviews many basic algorithms found within the literature and assesses their performance in a very controlled state of scenario. To solve this kind of problems genetic algorithm approach uses for feature selection and parameters optimization. This work involves brief study and designing a framework that incorporates genetic algorithm with SVM for feature selection and classification on the training dataset. Diagnosis of diseases is very important and difficult task, and this study helps for finding technique that performs feature selection and parameters setting. The objective of this work is to research existing work for feature set and discussed technique for improving classification accuracy.

   Index Terms Feature selection, SVM, Diagnosis, Cancer detection, Genetic Algorithm.

   1. INTRODUCTION

      Many practical pattern classification tasks need learning an appropriate classification function that assigns a given input pattern, usually represented by a vector of attribute values to a finite set of categories. Feature selection is employed to identify a powerfully predictive subset of fields in the database and reduce the number of fields given to the mining method. By extracting the maximum amount of data from a given data set while using smallest number of features. It is able to save important computation time and build models that generalize much better for unseen data points. According to Yang and Honavar [1], the selection of features are used to represent patterns which are given to a classifier affects many pattern classification aspects, including the accuracy of the learned classification rule, the time required for learning a classification function, the number of examples required for learning, and the cost associated with features.

      Genetic algorithms can generate both optimal feature set and SVM parameters at the time. Our analysis objective is to optimize the parameters and feature set, without any lose in the SVM classification accuracy. The proposed technique performs feature selection and parameters setting in an evolutionary

      way. Feature set selection algorithms are often classified into two categories: the filter approach and the wrapper approach [2, 3]. The wrapper approach to feature set selection is used in paper because of accuracy. Within the literature, few algorithms are proposed for SVM feature selection [4, 5, 6, 7, 8]. Some other GA-based feature selection strategies were also proposed [9]. However, these papers only focused on feature selection, not on the parameters optimization for the SVM classifier. [10] Proposed a GA-based feature selection approach which used theoretical bounds on the generalization error for SVMs.

      SVM is used to classify data with totally different class labels by determining a set of support vectors which are members of the set of training inputs that define a hyper plane in the feature space. SVM offer a standard mechanism that same as the hyper plane surface to the training values by use of a kernel function. The user might choose a kernel function (e.g. linear, polynomial, or sigmoid) for the SVM throughout the training process that selects support vectors along the surface of this function.

      While using SVM, two issues are confronted: the way to opt the optimum input feature set for SVM, and the way to set effective kernel parameters. These two issues are crucial, because the appropriate kernel parameters influenced by selection of feature subset and vice versa [10]. Therefore, for getting the optimum feature set and SVM parameters should occur at the same time.

   2. FEATURE SELECTION

      Feature selection algorithms fall under two categories: feature ranking and set selection. Feature ranking eliminates all features that do not achieve an adequate score and rank the features by metric. Set selection searches for the optimal set from the set of possible features. Feature selection (known as set selection) is a method used in machine learning, wherein for application of learning algorithm subsets of the available features are selected from data. The most effective set contains the smallest range of dimensions that contributes to accuracy; one discards the remaining unimportant dimensions. This is a stage of preprocessing which is very important and is one of two ways by which curse of dimensionality is avoided (the other is feature extraction) [11]. These are two approaches:

      Forward selection: Begin with no variables and add them one by one, at each step adding the one that decreases the error,

      until any more addition does not considerably decrease the error.

      Backward selection: Begin with all the variables and eliminate them one by one, at each step removing the one that decreases the error (or will increase it slightly); until any more removal increase the error considerably.

      There are two main models for feature selections that are filtering and wrapper model [12]. The filtering approach filter features independently from the induction algorithm after receiving a set of features. The wrapper model searches for best feature subsets, and evaluate them by using n-fold cross- validation on the training data. With any other induction algorithm this scheme may be used in conjunction and which can be used for evaluating feature subsets on the validation set. The simple greedy algorithm can be used for searching feature subsets. Greedy algorithms such as backward elimination or forward selection, or more complex ones that can both delete and add features at each step.

      Since the wrapper model needs more computation, filtering is the common kind of feature selection. This is often very true within the domain of textual information retrieval, where use of bag-of-words model results large number of features. It was found that CHI, information gain (IG) and document frequency (DF) are most effective (reducing the feature set by 90-98% with no performance penalty, or small performance increase as a result of removal of noise). Contrary to a popular belief in data retrieval that common terms are less informative and document frequency that prefers frequent terms (except for stop words) was found to be quite effective for text categorization.

      Advantages of feature selection

      It reduces the dimensionality of the feature space, to limit the requirement of storage and increase speed of algorithm;

      • It removes the redundant, noisy or irrelevant data.

      • The running time of learning algorithms are speeding up by immediate effects of data analysis tasks.

      • Data quality improves.

      • The accuracy of the resulting model increases.

      • Feature set reduction, to avoid wasting of resources in the next round of information collection or throughout utilization.

      • Performance improvement to achieve in predictive accuracy.

      • Data understanding, to achieve information aboutprocess that generated the info or just visualizes the info.

        Feature selection is analogous to data preprocessing technique. It is approach which is used to identify subset of features which are mostly related with target model The aim of feature selection is to increase the level of accuracy, reduce dimensionality; shorter training time and enhances generalization by reducing over fitting. Feature selection techniques are a subset of general field of feature extraction. Feature extraction use to creates new features from functions of the original features, whereas feature selection returns a set of the options. Feature selection techniques return a subset of features. Feature selection is used in domain where

        there are few samples (or data points) and comparatively many features.

        Steps in an every Feature selection Method:

      • Invention Procedure: Manufacture candidate set from original feature set.

      • Estimation Function: Estimate the candidate set.

      • Evaluation: Compare with user outlined threshold value.

      • Verification Method: Check out whether or not the set is valid.

      Fig. 1. Steps involved in feature selection

   3. GENETIC ALGORITHM

      Genetic algorithms (GA), is a general adaptive optimization search methodology which supports an analogy of Darwinian natural selection and genetics biological systems, could be a promising alternative to standard heuristic search. GA work with a collection of candidate solutions referred to as a population. Based on the Darwinian principle of survival of the fittest, the GA gains the optimum solution when a series of repetitive computations are applied. GA generates successive populations of alternate solutions which are represented by chromosomes, i.e. an answer to the problem, till acceptable results are obtained related to the characteristics of exploitation and exploration search. GA will cope with large search areas efficiently, and therefore has less likelihood to induce local optimal solution than other algorithms.

      Fig. 2. Genetic crossover and mutation operation.

      A fitness function assesses the standard of a solution in the analysis step. The crossover and mutation functions are the units that impact the fitness value. For reproduction, chromosomes are selected by evaluating the fitness value. The fitter chromosomes have higher chance to be elected into the recombination pool using the roulette wheel or the tournament selection methods.

      Fig. 3. Evolutionary cycle.

      Fig. 2 illustrates the genetic operators of crossover and mutation. Crossover, the critical genetic operator that enables new solution regions within the search space to be explored, it is a random mechanism for exchanging genes between two chromosomes using the one or two points crossover, or homologue crossover. In mutation the genes could often be altered, i.e. in binary code genes change genes code from 0 to 1 or vice versa. Offspring replaces the previous population using the diversity replacement strategy and forms a replaced or new population in the next generation. The evolutionary {biological process} process operates several generations till termination conditions satisfy. Fig. 2 depicts the GA evolutionary

      {biological process} process mentioned above [13] Main steps of a GA:

      1. Construction of the first generation

      2. Selection

         While stopping criteria not met do

      3. Crossover

      4. Mutation

      5. Selection End

      1. Encoding solutions

         The first step to perform GA is that the encoding of solutions. Indeed, in GA, each and every potential solution totally represented by a numerical vector. The historical encoding is bit strings however real encoding has a lot of uses and has sure benefits. In our case, every solution must contain the labels of peaks and the corresponding thresholds values [14].

      2. Initial population

         Like in any step by step optimization drawback, the information of good beginning parameters advantages the convergence speed of the algorithmic. But such kind of information is rarely available. This leads to cover large part of solution space due to generation of random initial population.

         Therefore, to make the exploration of the solution space easier, a very heterogeneous initial population is suitable. In our case, parameters to be set are the following, for each of the Spop individual of the initial population: the number of peaks , the set of peaks used and corresponding thresholds of peaks. Moreover, a peak having massive amplitude among the various spectra is likely to be more interesting for discrimination [15]. Thats why the probability to pick out one peak within the initial population is proportional to range of its intensity values among all the spectra. Then, for each individual, the peaks are chosen according to their amplitude and random numbers of peaks are computed. This selection leads to disadvantage very small peaks which might be remaining noise.

      3. Fitness values

         As discrimination of the various spectra is the objective, the fitness values have taken into account the good classification rate, achieved by every potential solution. Moreover, generally, a committee created of various peaks is seemingly to perform a good discrimination than a small one. However, using of several decision stumps could lead to over fit the training set and lose generality [14, 15]. So, a parsimony term concerning the number of peaks, , in the committee is added. As there is no priori information of the optimum range of decisions stumps required, the parsimony term, (a) are going to be defined as a linear function of the committee size (small sizes are favored):

         (a) = a+ The fitness function is given by-

         Fitness = +c× (a)

         In practice, (the good classification rate) is the first goal of the optimization, therefore it'll be favored. As and (a) belong to [0, 1], c must be a real number in [0, 1].

      4. Selection step

         This step supported the fitness values. According the fitness value individuals are ranked and highest rank is given to the best one. Then, to keep one solution in the next generation, its probability is

         P (selecting kth stratified solution) = +×k Where and are chosen so

         Spop

         + × k = 1

         k=1

      5. Crossover step

         The objective of this step is to gather interesting features (peaks and thresholds) of several solutions in new individuals by making combination of the previously retained solutions. Its necessary to notice that this step is independent from the optimization, that's to say, a crossover can produce good and worse solutions equally. Only the selection step is used to eliminate bad solutions.

      6. Mutation step

      This step brings the necessary hazard for efficiently explore the solution space. Any point of this space can be reached, it is

      assured. Moreover, if a local optimum is obtained, then a too quick convergence to this local optimum will be avoided by mutation. The mutation rate (proportion of the solutions which will undergo mutation), m is defined for every generation. At the beginning it is set to a maximum value then decreases to allow convergence and it increases again to avoid local optima, finally. At the beginning of the algorithmic, usually having probability of one mutation per individual (in this stage, the proportion of the various mutations is more of important), hence it's set to 0.9 [14, 15]. Mutation consists in changing values in the vectors corresponding to the solutions that have been chosen for undergoing mutation. In our context, mutations will be divided into three types:

      Peak elimination: Randomly a peak is chosen and removed from solution (i.e. the committee),

      Peak addition: A new peak is chosen, added and then the optimal threshold is associated,

      Threshold relocation: Randomly one of the thresholds is removed and relaced by another one.

   4. SUPPORT VECTOR MACHINE

      Support Vector Machine is constructed on the structural risk minimization principle to seek a decision surface that may separate the data points into two categories with a maximal margin between them. The selection of the correct kernel function is the main challenge when using a SVM. It might have completely different forms like Radial Basis function (RBF) kernel and polynomial kernel. The advantage of the SVM is its capability of learning in sparse, high dimensional spaces with only a few training examples by minimizing the empirical error and the complexness of the classifier at same time. . WEKA uses the Sequential Minimal Optimization (SMO) algorithm for SVM. The Support Vector Machines (SVM) form a group of methods stemming from the structural risk minimization principle, with the linear support vector classifier as its most basic member. The SVM aims at creating a decision hyper plane that maximizes the margin, i.e., the distance from the hyper plane to the nearest examples from each of the classes This allows for formulating the classifier training as a constrained optimization problem. Importantly, the objective function is unimodal, contrary to e.g. neural networks, and thus can be optimized effectively to global optimum. In the simplest case, compounds from different classes can be separated by linear hyper plane; such hyper plane is defined solely by its nearest compounds from the training set. Such compounds are referred to as support vectors, giving the name to the whole method. In most cases, however, no linear separation is possible. To take account of this problem, slack variables are introduced. These variables are associated with the misclassified compounds and, in conjunction with the margin, are subject to optimization. Thus, even though the erroneous classification cannot be avoided, it is penalized. Since the misclassification of compounds strongly influences the decision hyper plane, the misclassified compounds also become support vectors.

      Support vectors and margins in linearly separable (a) and non-separable (b) problems. In non-separable case, negative

      margins are encountered and their magnitude is subject to optimization along with the magnitude of the positive margins.

      Originally developed by Vladimir Vapnik [16], Support Vector Machines (SVMs) are a machine learning technique for supervised classification that have gained both popularity and momentum.

      Support Vector Machines are linear classifiers based on the concept of decision planes that define decision boundaries. A decision plane is one that separates between a set of objects having different class memberships. Additionally, SVMs are a classification method that determines the maximum-margin hyperplane [17]. This terminology is defined shortly. It should be noted that Support Vector Machines compete with Neural Networks as tools for solving pattern recognition problems.

      Support Vector Machines learn their classification through a training data set of the form

      The l instances of the training data each contain an n- dimensional vector ~x that describes the features of that instance and a label y that classifies the instance as belonging to one of two categories, 1 or -1 (positive or negative) respectively. Given sufficient training examples the Support Vector Machine is then able to classify previously unseen examples (instances of data), those with no predefined label, into one of the two categories [18].

      In the case of basic linear classification a Support Vector Machine creates a maximummargin hyperplane that lies in a potentially transformed input space. Given binary choice training examples (labeled either positive or negative), a maximum-margin hyperplane divides the positive and

      negative examples, such that the distance from between the respective class (in this case positive or negative), to the hyperplane is maximized. This is termed maximising the margin.

      This can also be viewed from a geometric standpoint. The Support Vector Machine attempts to construct a decision surface that bisects Rn such that all instances belonging to the positive class appear on one side of the surface with all instances belonging to the negative class appearing on the other. See Figure 4. While this approach is not new to the classification field, where SVMs set themselves apart is in their implementation.

      Fig. 4. Margin between classes.

      In order to obtain a maximum margin between a class and the decision surface we must define a convex hull for that class and maximise the margin in respect to that hull.

      This is because the closest approach of a particular class to the decision surface may not be at a specific point, but a linear combination of points.

      Formally, a convex hull is defined as a set of points S in n dimensions that is the intersection of all convex sets containing S [19]. For N points p1, …, pN, the convex hull C is then given by the expression:

      As previously stated, Support Vector Machines are linear classifiers that construct decision surfaces (hyper planes) between the convex hulls of classes. However, through the use of kernel functions, SVMs can find hyper planes in an extended attribute space which is equivalent to finding a non- linear separating surface in the original attribute space. This allows non-linear classification.

   5. VARIOUS FEATURE SELECTION TECHNIQUES

      1. Selected options exploitation fuzzy modeling

         The paper [20] has presented a study of medical data processing and data mining that is involving the use of eleven feature selection strategies and three fuzzy modeling strategies; such strategies aren't all available in a commercial data processing and data mining package. The objective is to determine that which combination of feature selection and fuzzy modeling strategies has the best performance for a given dataset. Two medical datasets and one industrial dataset were tested with fivefold stratified cross-validation. All the combination of feature selection and fuzzy modeling strategies were applied.

      2. Feature selection methodology supported association rules

         The paper [21] a hybrid methodology for identification of erythemato-squamous diseases supported Association Rules (AR) and Neural Network (NN). Feature extraction is t the key for pattern recognition and classification. If the features aren't chosen well, the best classifier can perform poorly. A feature extractor should reduce the feature vector to a lower dimension that contains most of the useful data from the initial vector. So, AR is employed for reducing the dimension of erythemato- squamous diseases dataset and NN is employed for intelligent classification. The projected AR+NN system performance is compared with NN model. The dimension of input feature area is reduced from 33 to 24 by using AR. In testing stage, proposed system performances are evaluated by applying 3- fold cross validation methodology to the erythemato-squamous diseases dataset. The classification rate of proposed system is 98.61% for 24 inputs. This analysis demonstrate that the AR will be used for reducing the dimension of feature vector and

         proposed AR+NN model will be used for obtaining efficient automatic diagnostic systems for other diseases.

      3. Gene selection multi-view fitness function

         The paper [22] provided a CD-MFS algorithm which is based on memetic evolutionary idea that uses accurate set of fuzzy if-then rules that can classify gene expression data. It begins with low quality rules, and results in high quality rule set. This algorithm classifies cancerous and benign tumors efficiently and has acceptable accuracy. 14_Tumors cancer dataset was evaluated by our proposed algorithm and it compared with other classification systems. Results indicate that our CD-MFS outperforms many well-known and up to date classifications. Moreover, the paper suggests new reasoning method and Multi-View fitness functions inmemetic algorithms. The introduced Multi-View fitness functions classifying cancerous tumors from gene expression data by considering both local and global fuzzy rule strength. This work tend to additionally targeted on producing meaningful fuzzy rules from the memetic algorithmic, that are more interpretable for a medical expert. On the other hand, each kind of tumor is clearly distinguishable by if-then rules produced by the algorithm.

      4. Feature generation using genetic programming

         [23] This paper presents a genetic programming based methodology to classify diabetes data. To facilitate the selection of features and for evaluating the effectiveness of diabetes features various methodologies have been used in this research. By making combinations of selected features GP has been used to automate the process of generating new features. A variation of GP which is called GP with CPS and it has been used which performs better than the standard GP.GP improves the performance and it reduces the eight dimensions to single dimension. The new features which generated by GP are tested by KNN and SVM to evaluate the performance and the results demonstrate that GP generated features show significant improvement in performance as compared to the performance achieved by original diabetes features. In comparison with other methods present in the literature shows the superiority of the proposed method.

      5. Feature selection method based hybrid intelligent system

      Ovarian cancer diagnosis is a vital study because early detection and accuracy staging are the keys to increase the survival rate of the patient. In papers, [24] propose a novel hybrid intelligent system, that derives simple yet convincing fuzzy inference rules to diagnose ovarian cancer and determine its stage according to the level of seriousness. Our proposed self organizing model is known as Genetic algorithmic and Rough Set Incorporated Neural Fuzzy System (GARSINFIS) that utilizes the inference rule base automatically derived by our proposed Genetic algorithmic program based Rough Set clustering (GARSC) technique. We combine the advantages of the individuals and alleviate certain limitation, by fusing various soft computing techniques together. Hospital data should be collected as real world data for applying GARSINFIS, two established medical data sets are

      benchmarked against other established models that focus on the compactness of the derived inference rules. All experimental results are encouraging, especially ovarian cancer diagnoses. As a result GARSINFIS requires limited number of constraints and control parameters. It needs no human intervention and expert guidance to achieve correct diagnoses when benchmarked against other models. Most important, it automatically derives rules and select features that are applied and suggested by doctors.

      Author	Year	Model	Processing techniques	Application
      Sean N. Ghazavi, [20]	2008	Fuzzy modeling	Fuzzy k-nearest neighbor, fuzzy clustering-based modeling, and fuzzy inference system	Diagnosis of breast cancer dataset
      Murat Karabataka [21]	2009	Association Rules	Feature selection method based on Association Rules (AR) and Neural Network (NN)	Diagnosis of erythemato- squamous diseases
      A. Zibakhsh [22]	2013	Memetic algorithm with a multi- view fitness function	Evaluates each single fuzzy if then rule according to the specified rule quality	Cancer tumor detection
      Muhammad Waqar Aslama [23]	2013	Genetic programming	Features selection using t- test, F-score selection, and genetic programming	Diabetes classification
      Di Wanga [24]	2014	Hybrid intelligent system	Self-organizing neural fuzzy inference system	Ovarian cancer diagnosis

      TABLE 1 COMPARISON OF EXISTING TECHNIQUES

   6. CONCLUSION

It is necessary for cancer patient to detect disease at very initial stage to survive and for better treatment. For this purpose feature selection play important role for detecting cancer at initial stage to diagnose at correct time. Feature selection process involves particular relation definition for different feature how they lead responsible for cancer. Thus Feature selection Algorithm (FSA) should be economic with respect to time and cost of detection and also need to be reliable as well. Lots of algorithms are given for this purpose which is based on totally different methods to explore various possibilities of finding rule as relation of different features with cancer disease. Feature selection technique is used to improve accuracy of classifier, reduce dataset and remove irrelevant data. This work gives comparative analysis of various existing feature selection methods and algorithms. This work also

discussed advantages with their drawbacks as well of feature selection strategies.

REFERENCES

1. Yang, J., & Honavar, V. (1998). Feature subset selection using a genetic algorithm.IEEE Intelligent Systems, 13(2), 4449.

2. John, G., Kohavi, R., & Peger, K. (1994). Irrelevant features and the subset selection problem. Proceedings of the 11th international conference on machine learning, San Mateo, CApp. 121129.

3. Kohavi, R., & John, G. (1997). Wrappers for feature subset selection.Artificial Intelligence, 97(12), 273324.

4. Bradley, P. S., & Mangasarian, O. L. (1998). Feature selection via concave minimization and support vector machines. In Proceedings of the 13th international conference on machine learning (pp. 8290). San Francisco, CA.

5. Bradley, P. S., Mangasarian, O. L., & Street, W. N. (1998). Feature selection via mathematical programming. INFORMS Journal on Computing, 10, 209217

6. Weston, J., Mukherjee, S., Chapelle, O., Pontil, M., Poggio, T., & Vapnik, V. (2001). Feature selection for SVM. In S. A. Solla,

   T. K. Leen, & K.-R. Muller, Advances in neural information processing systems(Vol. 13) (pp.668674). Cambridge, MA: MIT Press.

7. Guyon, I., Weston, J., Barnhill, S., & Bapnik, V. (2002). Gene selection for cancer classification using support vector machines. Machine Learning, 46(13), 389422

8. Mao, K. Z. (2004). Feature subset selection for support vector machines through discriminative function pruning analysis. IEEE Transactions on Systems, Man, and Cybernetics, 34(1), 6067

9. Raymer, M. L., Punch, W. F., Goodman, E. D., Kuhn, L. A., & Jain, A. K. (2000). Dimensionality reduction using genetic algorithms. IEEE Transactions on Evolutionary Computation, 4(2), 164171.

10. Fro¨hlich, H., & Chapelle, O. (2003). Feature selection for support vector machines by means of genetic algorithms. Proceedings of the 15th IEEE international conference on tools with artificial intelligence, Sacramento, CA, US App. 142148.

11. Fabrizio Sebastiani, "Machine Learning in Automated Text Categorization", ACM Computing Surveys, Vol.34, No. 1, pp. 147, 2002.

12. YZhao and G Karypis, "Hierarchical clustering algorithms for document datasets", Data Mining and Knowledge Discovery, pp. 141-168, 2005.

13. Davis, L. (1991).Handbook of genetic algorithms. New York: Van Nostrand Reinhold.

14. Christelle Reynes, Robert Sabatier, Nicolas Molinari, Sylvain Lehmann," A new genetic algorithm in proteomics: Feature selection for SELDI-TOF data", in Elsevier Computational Statistics and Data Analysis 52 (2008) 43804394.

15. . Dash, H. Liu," Feature Selection for Classification", in Elsevier Intelligent Data Analysis 1 (1997) 131156.

16. Vladimir N. Vapnik. The nature of statistical learning theory. Springer-Verlag New York, Inc., New York, NY, USA, 1995.

17. George Quievryn. Mutplus – glossary.

18. Douglas Clarke. An Investigation into Applying Support Vector Machines to Pixel Classification in Image Processing. Honours

    Thesis, School of Computer Science and Software Engineering, Monash University, 2003.

19. Eric W. Weisstein. Convex hull. from mathworlda wolfram web resource, 1999-2005.

20. Sean N. Ghazavi, Thunshun W. Liao, Medical data mining by fuzzy modeling with selected features, Artificial Intelligence in Medicine 43, Pages 195206, Elsevier, 2008.

21. Murat Karabataka, M. Cevdet Ince, A new feature selection method based on association rules for diagnosis of erythemato- squamous diseases, Expert Systems with Applications, 36 Pages 1250012505, Elsevier, 2009

22. A. Zibakhsh, M. Saniee Abadeh, Gene selection for cancer tumor detection using a novel memetic algorithm with a multi- view fitness function, Engineering Applications of Artificial Intelligence 26, Pages 12741281, Elsevier, 2013.

23. Muhammad Waqar Aslama, Zhechen Zhu, Asoke Kumar Nandi,

    Feature generation using genetic programming with comparative partner selection for diabetes classification, Expert Systems with Applications 40, Pages 54025412, Elsevier, 2013

24. Di Wanga, Chai Queka, Geok See Ng, Ovarian cancer diagnosis using a hybrid intelligent system with simple yet convincing rules, Applied Soft Computing 20, Pages 2539, Elsevier, 2014.