Setiono,R.; "Extracting rules from pruned neural networks for breast cancer diagnosis " Artificial Intelligence in Medicine 8(1):37-51, 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....Keywords Pareto optimization, di#erential evolution, artificial neural networks, breast cancer. 1 Introduction The economic and social values of Breast Cancer Diagnosis (BCD) are very high. As a result, the problem has attracted many researchers in the area of computational intelligence recently [6, 8, 10, 22, 26, 32, 33, 34]. Because of the importance of achieving highly accurate classification, Artificial Neural Networks (ANNs) are among the most common methods for BCD. Research in the area of using ANNs for medical purposes more specifically BCD [6, 8, 10, 22, 26, 32, 34] has been at the center of attention ....

....intelligence recently [6, 8, 10, 22, 26, 32, 33, 34] Because of the importance of achieving highly accurate classification, Artificial Neural Networks (ANNs) are among the most common methods for BCD. Research in the area of using ANNs for medical purposes more specifically BCD [6, 8, 10, 22, 26, 32, 34] has been at the center of attention for several years. Unfortunately, to our present knowledge, none of this type of research was able to enter the clinic either in terms of routine use or to replace the radiologist. This could be ascribed to a number of factors. The first problem was the ....

[Article contains additional citation context not shown here]

R. Setiono. Extracting rules from pruned neural networks for breast cancer diagnosis. Artificial Intelligence in Medicine, 8:37--51, 1996.

....little understandability, i.e. diagnostic decisions are essentially black boxes, with no explanation as to how they were attained. With increased interpretability in mind as a prior objective, a number of researchers have applied the method of extracting Boolean rules from neural networks [160 162, 169]. Their results are encouraging, exhibiting both good performance and a reduced number of rules and relevant input variables. Nevertheless, these systems use Boolean rules and are not capable of furnishing the user with a measure of confidence for the decision made. My own work on the evolution of ....

....2. 4) 125, 130] and with the systems obtained by Setiono s NeuroRule approach [161] note that the results presented by these two works were the best reported to date for genetic fuzzy and neuro Boolean rule systems, respectively, and that they were compared with other previous approaches such as [160,162,169]) The evolved 94 95 96 97 98 99 0 40 80 120 160 Classification performance Number of systems Figure 4.1 Summary of results of 495 evolutionary runs. The histogram depicts the number of systems exhibiting a given performance level at the end of the evolutionary run. The performance ....

R. Setiono. Extracting rules from pruned neural networks for breast cancer diagnosis. Artificial Intelligence in Medicine,8:37--51, 1996.

....each entry representing the classification for a certain ensemble of measured values: There are several studies based on this database. Among them, researchers having interpretability of the diagnostic as a prior objective, have applied the method of extracting Boolean rules from neural networks [40], 42] 43] Our own work on the evolution of fuzzy rules for the WBCD problem has shown that it is possible to obtain diagnostic systems exhibiting high performance, coupled with interpretability and a confidence measure [28] 30] In our previous work we used a simple genetic algorithm rather ....

....It exhibits an overall classification rate of 98.98 and its longest rule includes five variables. sults presented by these two works were the best reported to date for genetic fuzzy and neuro Boolean rule systems, respectively, and that they were compared with other previous approaches such as [40], 42] 43] The evolved fuzzy systems described in this paper can be seen to surpass those obtained by other approaches in terms of performance, while still containing simple, interpretable rules. As shown in Table IV, we obtained higher performance systems for all rule base sizes but one, ....

R. Setiono, "Extracting rules from pruned neural networks for breast cancer diagnosis," Artif. Intell. Medicine, vol. 8, pp. 37--51, 1996.

....exhibits little understandability, i.e. diagnostic decisions are essentially black boxes, with no explanation as to how they were attained. With increased interpretability in mind as a prior objective, a number of researchers have applied the method of extracting Boolean rules from neural networks [27,28,30]. Their results are encouraging, exhibiting both good performance and a reduced number of rules and relevant input variables. Nevertheless, these systems use Boolean rules and are not capable of furnishing the user with a measure of confidence for the decision made. Our preliminary work on the ....

....models represents valuable information to be used for our choice of fuzzy parameters. When defining our setup we took into consideration the following results, described in previous works: # Small number of rules. Systems with no more than four rules have been shown to obtain high performance [25,27]. # Small number of #ariables. Rules with no more than four antecedents have proven adequate [25,28,30] # Monotonicity of the input #ariables. Simple observation of the input and output spaces shows that higher valued variables are associated with malignancy. Some fuzzy models forgo ....

[Article contains additional citation context not shown here]

Setiono R. Extracting rules from pruned neural networks for breast cancer diagnosis. Artificial Intelligence in Medicine 1996:37 -- 51.

....little understandability, i.e. diagnostic decisions are essentially black boxes, with no explanation as to how they were attained. With increased interpretability in mind as a prior objective, a number of researchers have applied the method of extracting Boolean rules from neural networks [24 26]. Their results are encouraging, exhibiting both good performance and a reduced number of rules and relevant input variables. Nevertheless, these systems use Boolean rules and are not capable of furnishing the user with a measure of confidence for the decision made. Our preliminary work on the ....

....represents valuable information to be used for our choice 11 of fuzzy parameters. When defining our setup we took into consideration the following results, described in previous works: Small number of rules. Systems with no more than four rules have been shown to obtain high performance [22, 24]. Small number of variables. Rules with no more than four antecedents have proven adequate [22, 25, 26] Monotonicity of the input variables. Simple observation of the input and output spaces shows that higher valued variables are associated with malignancy. Some fuzzy models forgo ....

[Article contains additional citation context not shown here]

R. Setiono. Extracting rules from pruned neural networks for breast cancer diagnosis. Artificial Intelligence in Medicine, pages 37--51, 1996.

....little understandability, i.e. diagnostic decisions are essentially black boxes, with no explanation as to how they were attained. With increased interpretability in mind as a prime objective, a number of researchers have applied the method of extracting Boolean rules from neural networks [6 8]. Their results are encouraging, exhibiting both good performance and a reduced number of rules and relevant input variables. Nevertheless, these systems use Boolean rules and are not capable of furnishing the user with a measure of confidence for the decision made. Our preliminary work on the ....

....knowledge about the WBCD problem and about some of the extant rule based models represents valuable information to be used for our choice of fuzzy parameters. When defining our setup we took into consideration the following three results, described in previous works: 1) small number of rules [6,9]; 2) small number of variables [7 9] and (3) monotonicity of the input variables [9] Some fuzzy models forgo interpretability in the interest of improved performance. Where medical diagnosis is concerned, interpretability also called linguistic integrity is the major advantage of fuzzy ....

[Article contains additional citation context not shown here]

R. Setiono, "Extracting rules from pruned neural networks for breast cancer diagnosis," Artificial Intelligence in Medicine, pp. 37--51, 1996.

....It exhibits a classification rate of 99.33 , and an average of 2.3 variables per rule. the results presented by these two works were the best reported at the time for genetic fuzzy and neuro Boolean rule systems, respectively, and that they were compared with other previous approaches such as [43, 45, 47]) The evolved fuzzy systems described herein surpass those obtained by other approaches in terms of performance, while still containing simple, interpretable rules. As shown in Table 8, we obtained higher performance systems for all rule base sizes but one, i.e. from two rule systems all the way ....

R. Setiono. Extracting rules from pruned neural networks for breast cancer diagnosis. Artificial Intelligence in Medicine, 8:37--51, 1996.

....diagnostic decisions are essentially black boxes, with no explanation as to how they were attained. Kermani et al. 20] used a genetic algorithm to extract the most important variables, their attained performance level being lower (94.7 on all cases, no training test data was given) Setiono [21] proposed a method based on pruned neural networks for finding a set of rules to explain the diagnostic. His results are encouraging, exhibiting both good performance, and a reduced number of rules and relevant input variables. However, the extraction of rules is a manual, experience based ....

....all membership functions at any point is one. P and d define, the start point and the length of membership function edges, respectively (as shown above) search space size. Two output values are used, corresponding to Benign and Malignant diagnostics. ffl Number of rules: results from Setiono [21] show that few rules are needed to achieve good performance. Thus, we limited the number of rules to be in the range [1,4] These rules are evolved. 3. Connection parameters ffl Antecedents of rules: to be found by evolution. ffl Consequent of rules: the implemented strategy lets the algorithm ....

[Article contains additional citation context not shown here]

R. Setiono, "Extracting rules from pruned neural networks for breast cancer diagnosis," Artificial Intelligence in Medicine, pp. 37--51, 1996. 20

....5 1 1 : 1 Benign 2 5 4 4 : 1 Benign : 683 4 8 8 : 1 Malignant There are several studies based on this database. Among them, researchers having interpretability of the diagnostic as a prior objective, have applied the method of extracting Boolean rules from neural networks [36, 38, 39]. Our own work on the evolution of fuzzy rules for the WBCD problem has shown that it is possible to obtain diagnostic systems exhibiting high performance, coupled with interpretability and a confidence measure [24 26] In our previous work we used a simple genetic algorithm rather than Fuzzy ....

.... work [26] and with the systems obtained by Setiono s NeuroRule approach [37] note that the results presented by these two works were the best reported to date for genetic fuzzy and neuro Boolean rule systems, respectively, and that they were compared with other previous approaches such as [36,38,39]) The evolved fuzzy systems described in this paper can be seen to surpass those obtained by other approaches in terms of performance, while still containing simple, interpretable rules. As shown in Table IV, we obtained higher performance systems for all rule base sizes but one, i.e. from ....

R. Setiono. Extracting rules from pruned neural networks for breast cancer diagnosis. Artificial Intelligence in Medicine, 8:37-- 51, 1996.

....diagnostic decisions are essentially black boxes, with no explanation as to how they were attained. Kermani et al. 11] used a genetic algorithm to extract the most important variables, their attained performance level being lower (94.7 on all cases, no training test data was given) Setiono [12] proposed a method based on pruned neural networks for finding a set of rules to explain the diagnostic. His results are encouraging, exhibiting both good performance, and a reduced number of rules and relevant input variables. However, the extraction of rules is a manual, experience based ....

....variables (see Figure 1) We also experimented with three membership functions but the results were less satisfactory, probably due in part to the increased search space size. Two output values are used, corresponding to Benign and Malignant diagnostics. ffl Number of rules: results from Setiono [12] show that few rules are needed to achieve good performance. Thus, we limited the number of rules to be in the range [1,4] These rules are evolved. 3. Connection parameters ffl Antecedents of rules: to be found by evolution. ffl Consequent of rules: the implemented strategy has the algorithm ....

[Article contains additional citation context not shown here]

R. Setiono, "Extracting rules from pruned neural networks for breast cancer diagnosis," Artificial Intelligence in Medicine, pp. 37--51, 1996.

....5 1 1 : 1 Benign 2 5 4 4 : 1 Benign : 683 4 8 8 : 1 Malignant There are several studies based on this database. Among them, researchers having interpretability of the diagnostic as a prior objective, have applied the method of extracting Boolean rules from neural networks [18 20]. Our own work on the evolution of fuzzy rules for the WBCD problem showed that it is possible to obtain diagnostic systems exhibiting high performance, coupled with interpretability and a confidence measure [11 13] In our previous work we used a simple genetic algorithm rather than Fuzzy CoCo. ....

R. Setiono. Extracting rules from pruned neural networks for breast cancer diagnosis. Artificial Intelligence in Medicine, pages 37--51, 1996.

....the breast cancer database, the rules extracted by the Full RE from a simple MLP architecture (6 input, 6 hidden, and 2 output nodes) are presented in Table 7. The rules extracted by NeuroRule from the best among the pruned 100 MLP network architectures (6 inputs, 1 hidden, and 2 output nodes) are [36, 34]: Rule 1: If X 1 7:0 and X 2 8:0 and X 3 3:0 and X 8 9:0, then Benign Rule 2: If X 1 7:0 and X 2 8:0 and X 3 3:0 and X 6 9:0, then Benign Rule 3: If X 2 8:0 and X 3 3:0 and X 6 3:0 and X 8 9:0, then Benign Rule 4: Default Rule (Malignant) The corresponding rules extracted ....

R. Setiono. Extracting rules from pruned neural networks for breast cancer diagnosis. Artificial Intelligence in Medicine, 8(1):37--51, February 1996.

....are provided. Typically, one is interested in the generalization performance of the classifier, as estimated by dividing the instances into training, validation and or testing sets. Recently, some rule extraction strategies have also been applied to neural networks trained on this data set [39, 41]. In practice, one would like not only to generalize well and thus label future samples with reasonable competence, but also to be able to interpret the data and characterize it in a symbolic fashion. In this paper, we present a hybrid connectionist symbolic approach that addresses these twin ....

....extracted by our three techniques from a simple MLP architectures (6 input, 6 hidden, and 2 output nodes) are presented in Table 1, 2, and 3 respectively. The rules extracted by NeuroRule from the best among the pruned 100 MLP network architectures (6 inputs, 1 hidden, and 2 output nodes) are [41, 39]: R 1 : If X 1 7:0 and X 2 8:0 and X 3 3:0 and X 8 9:0, then Benign R 2 : If X 1 7:0 and X 2 8:0 and X 3 3:0 and X 6 9:0, then Benign R 3 : If X 2 8:0 and X 3 3:0 and X 6 3:0 and X 8 9:0, then Benign R 4 : Default Rule (Malignant) The corresponding rules extracted by ....

R. Setiono. Extracting rules from pruned neural networks for breast cancer diagnosis. Artificial Intelligence in Medicine, 8(1):37--51, February 1996.

....the breast cancer database, the rules extracted by the Full RE from a simple MLP architecture (6 input, 6 hidden, and 2 output nodes) are presented in Table 7. The rules extracted by NeuroRule from the best among the pruned 100 MLP network architectures (6 inputs, 1 hidden, and 2 output nodes) are [43, 41]: Rule 1: If X 1 7:0 and X 2 8:0 and X 3 3:0 and X 8 9:0, then Benign Rule 2: If X 1 7:0 and X 2 8:0 and X 3 3:0 and X 6 9:0, then Benign Rule 3: If X 2 8:0 and X 3 3:0 and X 6 3:0 and X 8 9:0, then Benign Rule 4: Default Rule (Malignant) The corresponding rules extracted ....

R. Setiono. Extracting rules from pruned neural networks for breast cancer diagnosis. Artificial Intelligence in Medicine, pages 37--51, February 1996.

....data set. Keywords: Neural networks, network pruning, rule extraction, hepatobiliary disorders, NeuroRule, NeuroLinear. 2 1 Introduction The e#ectiveness of artificial neural networks as tools that aid human decision making in the medical field has been reported in many recent papers [3, 4, 6, 11, 12, 14]. Experimental results indicate that neural networks perform particularly well in solving complex pattern classification problems due to their ability to model nonlinear relationship. Neural networks are also robust in handling data with noise or missing values due to their inherently parallel ....

....6 has been reached, redundant network connections are identified for removal. Connections with su#ciently small magnitude can be removed without a#ecting the classification accuracy of the network. Criteria for finding such connections have been developed and presented in our earlier papers [14, 19]. 3 Neural networks for diagnosing hepatobiliary disorders TABLE 1 HERE Each record in the database consists of the patient s sex and the results of nine biochemical tests for hepatobiliary disorders. The measurements obtained from these tests are Glutamic Oxalacetic Transaminase (GOT) ....

R. Setiono, Extracting rules from pruned neural networks for breast cancer diagnosis, Artificial Intelligence in Medicine 8(1) (1996) 37--51.

....of the input attributes of the data. A more concise set of rules can be thus expected from a network with fewer connections and fewer clusters of hidden unit activations. An application of NeuroRule to the Wisconsin breast cancer diagnosis (WBCD) problem has been reported in our previous work [14]. Since then, there have been a number of papers that introduce new methods for rule generation and their application to WBCD. Among the methods presented in these papers are a combined fuzzy genetic approach [11] and three neural network rule extraction algorithms [18, 19] A common feature of ....

....as needed to achieve the minimum required accuracy. In this paper, as in our previous work, we adopt the first approach. The number of hidden units needed for classifying the samples of the WBCD data set using a single hidden layer feedforward neural networks is as few as three and as many as nine [4, 14, 19]. Network connections link units in the input layer to units in the hidden layer and units in the hidden layer to those in the output layer. There is no direct connections between units in the input layer and units in the output layer. Given this network structure, it is natural to decompose the ....

[Article contains additional citation context not shown here]

R. Setiono, Extracting rules from pruned neural networks for breast cancer diagnosis, Artificial Intelligence In Medicine 8(1) (1996) 37--51.

No context found.

Setiono,R.; "Extracting rules from pruned neural networks for breast cancer diagnosis " Artificial Intelligence in Medicine 8(1):37-51, 1996.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC