Jsoftware.us

JOURNAL OF SOFTWARE, VOL. 6, NO. 12, DECEMBER 2011 Mining a Small Medical Data Set by Integrating the Decision Tree and t-test Department of Obstetrics and Gynecology, Chang Gung Memorial Hospital, Taipei, Taiwan 25137, R.O.C. Email: [email protected] Chien-Chou Shih2,Ding-An Chiang1 and Chun-Chi Chen1 * 1 Department of Computer Science & Information Engineering, Tamkang University, Tamsui, Taipei County, Taiwan 2 Department of Information & Communication, Tamkang University, Tamsui, Taipei County, Taiwan 25137, R.O.C. Email: [email protected], [email protected], [email protected] Abstract—Although several researchers have used statistical
In recent years, several researchers and our team have methods to prove that aspiration followed by the injection of
used statistical methods to prove that aspiration followed 95% ethanol left in situ (retention) is an effective treatment
by the injection of 95% ethanol left in situ (retention) is for ovarian endometriomas, very few discuss the different
an effective treatment for ovarian endometriomas [11]. conditions that could generate different recovery rates for
They always divide all patients into two groups group 1 the patients. Therefore, this study adopts the statistical
(ethanol irrigation) and group 2 (ethanol retention) method and decision tree techniques together to analyze the
postoperative status of ovarian endometriosis patients under
regardless of other conditions, such as the size of the cyst. different conditions. Since our collected data set is small,
In endometriomas-like dataset, the analysis of the containing only 212 records, we use all of these data as the
influence of treatment effectiveness use statistical and t- training data. Therefore, instead of using a resultant tree to
test cannot integration data mining in related research generate rules directly, we use the value of each node as a
[12-16] such as t-test [12], Logistic regression [15, 17], cut point to generate all possible rules from the tree first.
Decision trees [18-20], and SVM [15, 17]. However, Then, using t-test, we verify the rules to discover some useful
different conditions could generate different recovery description rules after all possible rules from the tree have
rates for the patients, and very few researchers discuss been generated. Experimental results show that our
this situation. Therefore, our aim is to investigate approach can find some new interesting knowledge about
recurrent ovarian endometriomas under different recurrent ovarian endometriomas under different
conditions.
conditions. We adopt the statistical method and decision tree techniques together to analyze the postoperative Index Terms—Data mining, Decision tree, t-test, p-value,
status of ovarian endometriosis patients. Ovarian endometriomas
The use of machine learning algorithms for the building of predictive and descriptive data mining models has become widely accepted in medical applications. Various models including Decision trees, Decision rules, The use of classification algorithms in the medical Logistic regression, Artificial neural networks, and SVMs domains has increasingly been the object of study in have been tested in a wide variety of clinical and medical recent years [1-3]. Although many conventional applications [21]. In order to resolve the endometriomas treatments are available in clinical medicine for patients problem, we should identify and treat the cause of the suffering from endometriosis, it is a common disease problem correctly. But such correct diagnosis and among women of reproductive age with a high recurrence treatment require the patients to have extensive live, rate, regardless of the treatment type [4]. In recent years, womb, uterocervical canal, laparoscopy, and others [22]. ultrasound-guided aspiration combined with drug therapy Computationally, there have been attempts from the has become a new alternative for patients since the endometriomas medical domain to analyze various ultrasound-guided aspiration of ovarian endometriomas treatment factors to predict the success of therapy. was proposed in 1991 [5]. Therefore, to reduce the high Previous studies that have utilized t-test for endometrial recurrence rate, some medical treatments have combined related research have only used this technology to ultrasound-guided aspiration with tetracycline [6], determine the effectiveness of treatment [23]. Based on methotrexate [7], recombinant interleukin-2 [8], or the decision-tree analysis, the optimal rule to detect the ethanol [9-11]. ultrasound characteristics of endometriomas in pre- and postmenopausal patients and to develop rules that characterize endometriomas [20]. For ovarian tumors * Corresponding author. and pregnancies of unknown location on medical 2011 ACADEMY PUBLISHER JOURNAL OF SOFTWARE, VOL. 6, NO. 12, DECEMBER 2011 decision making (prediction), mathematical algorithms II. MATERIALS AND BACKGROUND KNOWLEDGE are applied to data sets in order to obtain a model [15, 16]. Kinkel [24] et al. demonstrated in a meta-analysis of A. Patients indeterminate masses in sonography the superiority of This study was approved by the Institutional Review MRT over CT and Doppler sonography in predicting Board of our hospital. It was a retrospective review of malignancy. Fewer models have been hitherto developed, 212 consecutive patients treated at the outpatient even though the reliable identification of borderline and gynecological department of Chang Gung Memorial ovarian endometriomas would be a good step forward for Hospital, Taipei, Taiwan from July 1994 to July 2008. clinical practice. All patients had undergone previous surgical treatment The decision tree in classification algorithms has been for ovarian endometriomas and were being seen because applied to categorical attributes and numeric attributes in of a recurrence. Recurrence was defined as when one or different domains [25]. Since medical data always more persistent pelvic cysts greater than 3.0 cm were contain numeric attributes and handling them is a critical detected in two consecutive ultrasonographic task in inductive learning, this task has already been examinations. Transvaginal ultrasound-guided cyst embedded within the decision tree algorithm. Therefore, aspiration and ethanol injection were done on an outpatient basis. Patients randomly received ethanol we adopt the decision tree technique to analyze instillation at 0 min (ethanol injected, then immediately postoperative status of ovarian endometriosis patients in removed), 3–9 min, 10 min, or retention. The procedure this study. The decision tree is built by performing a was in accordance with that reported previously [8, 23] . heuristic-based local search to select the best test attribute The medical records include basic information of as the root of the decision tree. So, decision tree creates a patients, treatment-related information, and clinical branch for each value of that appearing in the training examination data obtained from the first visit and three- data. Then, the same procedure is operated on each month, six-month to one year follows-up. The medical branch to induce the remaining levels of the decision tree data set has 57 attributes, such as age, number of previous until all examples in a leaf belong to the same class. pregnancies, degree of menstrual cycle pains, urge Whereas the tree is represented by a set of rules, each technology to help the patient get pregnant, size of the branch of the decision tree is only represented by a rule. cyst, CA125 blood test value, types of surgery, ethanol However, since each branch of the decision tree is only irrigation duration. Usually this raw data contained lots of represented by a rule and the resultant rules of the tree are null values (missing information), which might have local, some useful description rules cannot be found by affected the accuracy of the study. In clinical practice the tree-generation algorithm. Therefore, instead of using data collection is difficult because the value of attention a resultant tree to generate classification rules directly, we to patient privacy; so data and collection methods are use the value of each node as a cut point to generate all limited. We required a preprocessing procedure to possible rules from the tree. We can filter out useless prepare the valid data for the analysis. rules by using the method of setting minimum recovery B. Decision Trees rates. However, in the study case, our collected data set is A decision tree is built by selecting the best test small, containing only 212 records, and we use all of attribute as the root of the decision tree. Then, the same these data as the training data. So because there are procedure is operated on each branch to induce the different conditions in a possible rule that could generate remaining levels of the decision tree until all examples in different recovery rates for the patients and very few a leaf belong to the same class. Decision trees can be researchers discuss this situation, our aim is to investigate categorized by data processing functions into the recurrent ovarian endometriomas under the different classification trees and regression trees. The common conditions in possible rule. Under this mining goal, we decision tree algorithms are compared in Table I. A use t-test to verify rules to discover some useful classification tree is applied to discrete variables, while a description rules after all possible rules from the tree are regression tree is applied to quantitative variables. The regression tree was first brought up by Breiman [26] in The models were built in collaboration with the introduction of Classification and Regression Trees gynecologists, and resulted in accurate predictions. The ovarian endometriomas models were based on multi-center data and have successfully passed prospective internal and external evaluations. The models for COMPARISON OF DECISION TREE ALGORITHMS pregnancy of unknown location were based on single Name center data and have passed the first internal evaluation. However, a large multi-center study is ongoing, aiming ID3 Discrete Classification for a thorough validation and for the development of new C4.5 Numeric Regression models for pregnancies of unknown location. CHAID Class Classification Discrete and Classification and 2011 ACADEMY PUBLISHER JOURNAL OF SOFTWARE, VOL. 6, NO. 12, DECEMBER 2011 (CART). Usually, in CART analysis, data is categorized strengthen our conclusion, we still divided all the patients into quantitative and discrete data. Quantitative data can are into two different groups according to the be applied to prediction, while discrete data can be corresponding rules obtained from the tree. The applied to classification. In other words, CART analysis significance level is 0.05. The recovery rate of these 212 is able to simultaneously process both quantitative and patients is 43.4%. discrete data. In the article, we use CART trees in the software of DB2 Intelligent Miner for Data, version 8.1 (IBM Corp., New York), to analyze the collected data to generate the resultant tree. Ⅲ. MINING USING DECISION TREE AND RESULTS In the article, we use the CART function in IBM DB2 Intelligent Miner for Data, version 8.1, to analyze the collected data to generate the resultant tree. The first step in reducing the dimensionality is to use PCA algorithm by using the dataset, and decision tree is used for the Figure 1. The resultant decision tree. second feature and rule extraction. After that in each iteration of the loop (step 1) confirmation attributes is selected into the output set (step 3). The selected As shown by test 1 in Table II, according to the rule classification attributes included Cyst_size, CA_125, and "Cyst_size ≥ 4.25 ==> Recovery = 0" all patients are BMI. Although the data set has many attributes, as shown divided into two groups by the cut point of the Cyst_size in figure 1, only four attributes are selected by the system attribute. Group 1 contains patients whose cyst size is less to build the decision tree. These four attributes are than 4.25, while cyst size of patients in group 2 are not Cyst_size, CA_125, BMI, and Recovery (Recovery less than 4.25. From test 1 in the table, we conclude that means no operation within six months and pregnant when all patients are divided into two groups: group 1 within six months or cyst size less than 3 cm in the sixth (Cyst_size <4.25) and group 2 (Cyst_size ≥ 4.25). The month). Therefore, when the patient has recovered, the recovery rate of group 1 (43/59, 72.88%) is significantly value of this attribute is set to 1; otherwise, it is set to 0. (p = 1.47E-08) greater than that of group 2 (49/153, 32%). Moreover, Cyst_size, CA_125, and BMI are the size of In other words, the recovery rate (or recurrence rate) is cyst, blood test value, and body mass index before affected by the cyst size regardless of the treatment type. operation, respectively. The resultant rules generated Moreover, from test 2 and 3 in Table Ⅱ, we also from the decision tree are shown as follows: conclude that the recovery rate of group 1 is significantly ≥ 4.25 Then Recovery = 0 (recovery better than that of group 2. Since each branch of the decision tree is only (2) If Cyst_size < 4.25 and CA_125 ≥ 146.81Then represented by a rule, and the resultant rules of the tree Recovery = 0 (recovery rate =0%) are local [27], some useful description rules cannot be (3) If Cyst_size < 4.25 and CA_125 < 146.81and found by the tree-generation algorithm. For example, the BMI < 22.7516 Then Recovery = 1 (recovery rate rule "Cyst_size < 4.25 and CA_125 ≥ 146.81 ==> =80.77%) Recovery = 0" should be ignored in the discussion (4) If Cyst_size < 4.25 and CA_125 < 146.81and because only three of the patients are related to this rule BMI ≥ 22.7516 Then Recovery = 0 (recovery rate =25%) and the over-fit problem. Consequently, the question is "for those patients whose Cyst_size < 4.25, whether the The CART is a binary tree, as shown in figure 1, there BMI value will affect the recovery rate." Since our data is a significant difference between two branches of a node size is very small and there are only three cut points in when the pruning process is performed automatically by this example, to overcome the above problems, we could the mining tool. Therefore, we can interpret the use these three cut points to generate all possible rules corresponding rules from the tree directly. However, to from the tree. We discuss this phenomenon in the COMPARISON OF DECISION TREE ALGORITHMS 2 Cyst_size <4.25 and CA_125 <146.81 Cyst_size <4.25 and CA_125 146.81 Cyst_size <4.25 and CA_125 <146.81 and Cyst_size <4.25 and CA_125 <146.81 and 2011 ACADEMY PUBLISHER JOURNAL OF SOFTWARE, VOL. 6, NO. 12, DECEMBER 2011 following section. according to three different conditions, as shown in Table III. As shown by test 1 in Table III, all patients are Ⅳ. INTERPRETING MINING RESULTS REGARDLESS TO divided into two groups according to the cut point of the CA_125 attribute. Group 1 contains patients whose CA_125 values are less than 146.81, while CA_125 From the tree in figure 1, we obtain the cut points of values of patients in group 2 are not less than 146.81. Cyst_size, CA_125, and BMI numeric attributes as 4.25, From test 1 in the table, we conclude that when all 146.81, and 22.752, respectively. We can use these three patients are divided into two groups, the recovery rate of values to generate all possible rules from the tree. group 1 (89/189, 47.09%) and group 2 (1/13, 7.69%) is Because the data set only contains 212 records, it is small. significantly different (p = 0.002769). From tests 2 and 3, So, we used all of these data, as in training data. However, we conclude that when a patient's CA_125 146.81, there there are different conditions for generating all possible is no significant difference between the recovery rates of rules that could generate different recovery rates for the the two groups. In other words, when CA_125   146.81, patients. Therefore, using single rule information, we the recovery rate (or recurrence rate) will mainly be cannot filter useful rules from all possible rules. And the affected by this CA_125 value regardless of the cyst size study case aim is to investigate recurrent ovarian and BMI value. That is, the recovery rate of that patient endometriomas under different conditions in a possible would be very low regardless of treatment type. These rule. For these reasons, we use t-test to verify rules to new useful descriptions cannot be discovered by the tree- discover some useful description rules after all possible generation algorithm directly. Actually, when a patient's rules from the tree are generated. Then, we use t-test to CA_125 ≥ 146.81, the patient must have other diseases; verify all possible rules obtained from the decision tree to otherwise, the value would not be greater than or equal to discover useful knowledge. Moreover, since the values of BMI and CA_125 may be missing, the number of patients of BMI or CA_125 attribute is not equal to 212. B. Effect of Cyst_size and BMI Values As indicated in the above section, when CA_125 ≥ A. Effect of CA_125 Value 146.81, the recovery rate (or recurrence rate) will mainly The rule "Cyst_size < 4.25 and CA_125 ≥ 146.81 ==> be affected by this CA_125 value regardless of the Cyst- Recovery = 0" should be ignored in the discussion _size and BMI values. Therefore, we do not consider this because of the over-fit problem. However, there are 13 attribute in the following discussion. According the cut patients whose CA_125 value is greater than 146.81 in points of the Cyst_size and BMI attributes, as shown in the whole training data. Therefore, we should check Table Ⅳ, all patients can be divided into two groups whether this CA_125 value presents some interesting according to four different conditions. From Table Ⅳ, we descriptions about the recovery rate. conclude that the recovery rate (or recurrence rate) will In this section, all patients are divided into two groups mainly be affected by the cyst size. RECOVERY RATES OF GROUPS 3 Cyst_size <4.25 and BMI< 22.7516 Cyst_size <4.25 and BMI 22.7516 4 Cyst_size  4.25 and BMI< 22.7516 Cyst_size 4.25 and BMI 22.7516 T-TEST RESULTS FOR CA_125 1 CA_125<146.81 2 CA_125 146.81 and Cyst_size<4.25 CA_125 146.81 and Cyst_size 4.25 3 CA_125 146.81 and BMI< 22.7516 CA_125 146.81 and BMI 22.7516 2011 ACADEMY PUBLISHER JOURNAL OF SOFTWARE, VOL. 6, NO. 12, DECEMBER 2011 From test 2 in Table II, we conclude that, for patients whose cyst size is less than 4.25 and CA_125 is not less RECOVERY RATES OF ETHANOL IRRIGATION AND ETHANOL RETENTION than 146.81, the BMI value, 22.7516, will significantly GROUPS ABOUT DIFFERENT CONDITIONS affect the recovery rate. However, we already know that when a patient's CA_125 is greater than 146.81, the recovery rate of the patient is very low. Therefore, the question in mind is that how about for those patients whose cyst size is less than 4.25. Will the value of BMI significantly affect the recovery rate of patients regardless of the CA_125 value? From test 3 in Table Ⅳ, we CA_125 <146.81 conclude that the recovery rates of group 1 (43/59, 72.88%) and group 2 (1/5, 20%) are significantly different (p = 0.002115). In other words, we can say that when the cyst size is less than 4.25, the BMI value will Cyst_size <4.25 affect the recovery rate. We get a new useful description that cannot be discovered by the tree-generation algorithm directly. This new description can be interpreted as follows: "For those patients whose cyst size BMI < 22.7516 is less than 4.25, when the patient's BMI value is less than 22.7516, the recovery rate is 78.72%; otherwise, the recovery rate is 20% only. That is, for those patients whose cyst size is less than 4.25, when the BMI value is Cyst_size <4.25 And BMI< not less than 22.7516, the value of BMI will significantly affect the recovery rate." Cyst_size <4.25 And C. Interpreting Mining Results with Ethanol Instillation Cyst_size 4.25 And BMI< As pointed out by Noma and Yoshida [9], ethanol instillation into the cyst cavity for more than 10 min was Cyst_size 4.25 And most effective at reducing the recurrence rate. Therefore, another goal is to verify whether aspiration followed by the injection of 95% ethanol left in situ is an effective found by our approach. However, our method works only treatment for ovarian endometriomas under different for a simple tree. To deal with more complex decision conditions. Therefore, in the data preprocessing step, we trees, we want to integrate other data mining the classify the values of ethanol irrigation duration into two approaches to analyze postoperative status of ovarian types: less than 10 min and retention. Therefore, for each endometriosis patients in the future. We hope to find condition, all patients are divided into two groups (group more precise description knowledge about the recovery 1: ethanol irrigation; group 2: ethanol retention) by the rates of the patients under different conditions. resultant cut points. The values of t-test for the above conditions are shown in Table V. From the table, we give the following conclusions: (1) When CA_125 > 146.81, recurrence rate is very [1] . Tang, T.I., et al., "A Comparative Study of Medical Data Classification Methods Based on Decision Tree high regardless of treatment type. and System Reconstruction Analysis," Industrial (2) When Cyst_size ≥ 4.25, the recovery rate of group Engineering and Management Systems, vol. 4, p. 1 (11/49, 22.45%) and group 2 (38/104, 36.54%) is 102¡V108, 2005. significantly different (p = 0.041180). Cui, Z. and G. Zhang, "A novel medical image (3) Except the above two conditions, although the dynamic fuzzy classification model based on ridgelet recovery rate of group 2 is better than that of group 1, transform," Journal of Software, vol. 5, pp. 458-465, there is no significant difference between these two Scheetz, L.J., J. Zhang, and J. Kolassa, "Classification tree modeling to identify severe and moderate vehicular injuries in young and middle-aged adults," Ⅴ. DISCUSSION AND CONCLUSION Artificial Intelligence in Medicine, vol. 45, pp. 1-10, Endometriosis is a complex disease with several attributes. Each of the attributes, such as the stage of the Ozkan, S. and A. Arici, "Advances in treatment options of endometriosis," Gynecol Obstet Invest, vol. disease, symptoms of the disease, serum level of CA-125, 67, pp. 81-91, 2009. can affect the making of treatment decisions and the [5] . Aboulghar, M.A., et al., "Ultrasonic transvaginal results of the treatment. aspiration of endometriotic cysts: an optional line of This study integrates the statistical method and treatment in selected cases of endometriosis," Human decision tree techniques to analyze the postoperative Reproduction, vol. 6, p. 1408, 1991. status of ovarian endometriosis patients. Experimental [6] . Aboulghar, M.A., et al., "Treatment of recurrent results show that some new interesting knowledge can be chocolate cysts by transvaginal aspiration and 2011 ACADEMY PUBLISHER JOURNAL OF SOFTWARE, VOL. 6, NO. 12, DECEMBER 2011 tetracycline sclerotherapy," Journal of assisted [22] . Kim, I.C. and Y.G. Jung, "Using Bayesian networks reproduction and genetics, vol. 10, pp. 531-533, 1993. to analyze medical data," in Lecture Notes in Artificial Mesogitis, S., et al., "Combined ultrasonographically Intelligence (Subseries of Lecture Notes in Computer guided drainage and methotrexate administration for Science), Leipzig, 2003, pp. 317-327. treatment of endometriotic cysts," The Lancet, vol. Hsieh, C.L., C.S. Shiau, and L.M. Lo, "Effectiveness 355, pp. 1160-1160, 2000. of ultrasound-guided aspiration and sclerotherapy [8] . Acien, P., et al., "GnRH analogues, transvaginal with 95% ethanol for treatment of recurrent ovarian ultrasound-guided drainage and intracystic injection endometriomas," Fertility and sterility, vol. 91, pp. of recombinant interleukin-2 in the treatment of 2709-2713, 2009. endometriosis," Gynecologic and obstetric [24] . Kinkel, K., et al., "Indeterminate ovarian mass at US: investigation, vol. 55, pp. 96-104, 2000. Incremental value of second imaging test for [9] . Noma, J. and N. Yoshida, "Efficacy of ethanol characterization-meta-analysis and Bayesian sclerotherapy for ovarian endometriomas," analysis," Radiology, vol. 236, pp. 85-94, 2005. International Journal of Gynecology & Obstetrics, vol. Wang, H. and P. Zhang, "A quantitative method for 72, pp. 35-39, 2001. pulse strength classification based on decision tree," Akamatsu, N., et al., "Ultrasonically Guided Puncture Journal of Software, vol. 4, pp. 323-330, 2009. of Endometrial Cyst: Aspiration of Contents and Breiman, L., et al., "Classi cation and regression Infusion of Ethanol," Acta Obstetrica et trees," Wadsworth, Belmont, 1984. Gynaecologica Japonica, vol. 40, pp. 1214-1215, Wang, K., S. Zhou, and Y. He, "Growing decision trees on support-less association rules," in Proceeding [11] . Hsieh, C.L., et al., "Effectiveness of ultrasound- of the Sixth ACM SIGKDD International Conference guided aspiration and sclerotherapy with 95% ethanol on Knowledge Discovery and Data Mining, Boston, for treatment of recurrent ovarian endometriomas," MA, 2000, pp. 265-269. Fertility and sterility, vol. 91, pp. 2709-2713, 2009. Akan, E., et al., "Predictive Power of Activin A Levels in the Prognosis of First Trimester In Vitro Fertilization Pregnancies," Journal of Women's Health, Ming-Yang Chang, M.D. is currently working as an
Associated professor in Gynecology, Chang Gung University School of Medicine, Taoyuan, Taiwan. His research interests Acien, P., et al., "Use of intraperitoneal interferon include Infertility, endometriosis and female reproductive [alpha]-2b therapy after conservative surgery for endometriosis and postoperative medical treatment with depot gonadotropin-releasing hormone analog: a randomized clinical trial* 1," Fertility and sterility, Chien-Chou Shih received the Ph.D. degree in information
engineering from Tamkang University, Taiwan, in 1998. He is vol. 78, pp. 705-711, 2002. currently an Assistant Professor with the Department of Takano, M., et al., "The impact of complete surgical Information and Communication, Tamkang University. His staging upon survival in early-stage ovarian clear cell research interests include embedded software programming, carcinoma a multi-institutional retrospective study," data mining applications, and engineering design education. International Journal of Gynecological Cancer, vol. 19, pp. 1353-1357, 2009. Dina-An Chiang received the BS degree in hydraulic
Van Calster, B., et al., "Towards a clinical decision engineering from Chung Yuan Christian University, Taiwan, in support system for pregnancies of unknown location," 1981, and the MS and PhD degrees in computer science from in Proceedings - IEEE Symposium on Computer- the University of Southwestern Louisiana in 1986 and 1990, Based Medical Systems, Jyvaskyla, 2008, pp. 581-583. respectively. He is currently a professor in the Department of Ben VAN CALSTE, "Predictive diagnostic models Computer Science. His research interests include fuzzy, for gynecologic applications with focus on multi-class relational databases and data mining. classification," PhD thesis, Dept. of Electrical Engineering, Katholieke Universiteit Leuven, Chun-Chi Chen received the MS degrees in computer
science and information engineering from Tamkang University [17] . Kyama, C.M., et al., "Evaluation of endometrial in Taipei, Taiwan, in 2003. His research interests include biomarkers for semi-invasive diagnosis of relational databases and data mining. endometriosis," Fertility and sterility, vol. 95, pp. 1338-1343.e3, 2011. Seeber, B., et al., "Panel of markers can accurately predict endometriosis in a subset of patients," Fertility and sterility, vol. 89, pp. 1073-1081, 2008. Liu, H., et al., "Detection of endometriosis with the use of plasma protein profiling by surface-enhanced laser desorption/ionization time-of-flight mass spectrometry," Fertility and sterility, vol. 87, pp. 988-990, 2007. [20] . Van Holsbeke, C., et al., "Endometriomas: Their ultrasound characteristics," Ultrasound in Obstetrics and Gynecology, vol. 35, pp. 730-740, 2010. Bellazzi, R. and B. Zupan, "Predictive data mining in clinical medicine: Current issues and guidelines," international journal of medical informatics, vol. 77, pp. 81-97, 2008. 2011 ACADEMY PUBLISHER

Source: http://www.jsoftware.us/vol6/jsw0612-24.pdf