Zdzislaw S. Hippe

Melanoma prediction using data mining system LERS

One of the important tools for early diagnosis of malignant melanoma is the total dermatoscopy score (TDS), computed using the ABCD (asymmetry, border, color, diameter) formula. Our primary objective was to check whether the ABCD formula is optimal. Using a data set containing 276 cases of melanoma and the LERS (Learning from Examples based on Rough Sets) data mining system, we checked more than 20,000 modified formulas for ABCD, computing the predicted error rate of melanoma diagnosis using 10-fold cross-validation for every modified formula. As a result, we found the optimal ABCD formula for our setup: discretization based on cluster analysis, the LEM2 (Learning from Examples Module, version 2) algorithm (one of the four LERS algorithms for rule induction) and the standard LERS classification scheme. The error rate for the standard ABCD formula was 10.21 %, while for the optimal ABCD formula the error rate was reduced to 6.04%. Some research in melanoma diagnosis shows that the use of the ABCD formula does not improve the error rate. Our research shows that the ABCD formula is useful, since, for our data set, the error rate without the use of the ABCD formula was higher (13.73%).

Optimization of the ABCD Formula Used for Melanoma Diagnosis

The ABCD formula is used for computing a new attribute, called TDS, to help with melanoma diagnosis. In our research four discretization techniques were used, two of them never published before. We found four corresponding new ABCD formulas to compute TDS by applying more than 163 thousand experiments of variable ten-fold cross validation. Diagnosis of melanoma with each of these new ABCD formulas, when used with an appropriate discretization technique, is significantly more accurate (with the level of significance 5%) than diagnosis using the traditional ABCD formula. Finally, the rule sets, induced from data sets obtained using four new ABCD formulas and the traditional ABCD formula, were graded by an experienced melanoma diagnostician.

A comparison of six approaches to discretization: a rough set perspective

We present results of extensive experiments performed on nine data sets with numerical attributes using six promising discretization methods. For every method and every data set 30 experiments of ten-fold cross validation were conducted and then means and sample standard deviations were computed. Our results show that for a specific data set it is essential to choose an appropriate discretization method since performance of discretization methods differ significantly. However, in general, among all of these discretization methods there is no statistically significant worst or best method. Thus, in practice, for a given data set the best discretization method should be selected individually.

Postprocessing of rule sets induced from a melanoma data set

The data mining system LERS (learning from examples based on rough sets) was used to induce rule sets from a data set describing melanoma (a dangerous skin cancer). The main objective of our research was to decrease the error rates for diagnosis of two fatal forms of melanoma based on these rule sets. The improvement was accomplished using two different techniques for postprocessing of rule sets: changing of rule strengths and rule truncation cutoffs. A rule strength is defined as the number of training cases correctly, classified by the rule. Rule truncation means an elimination of weaker rules. The criterion for the choice of the optimal form of the rule sets was the minimum of the sum of error rates for diagnosis of the two fatal forms of melanoma. Our research shows that at the cost of a minimal increase of the total error rate for patients that do not need immediate help, the sum of error rates for dangerous forms of melanoma may, be highly decreased. Also, for the optimal rule set, the sum of error rates for all forms of melanoma is minimal as well.

Data mining methods supporting diagnosis of melanoma

Melanoma, a dangerous skin cancer, is usually diagnosed using the ABCD formula. The main objective of our research was to find a better formula resembling the original ABCD formula using four different discretization methods. All four corresponding modified ABCD formulas are significantly more accurate (with the level of significance 5%) than the original ABCD formula. Our additional objective was to calibrate the rule set induced from the original data set, describing melanoma, using the best discretization method, so that the sensitivity (the conditional probability for recognition of malignant and suspicious melanoma) was increased. This objective was accomplished using a technique of changing rule strengths.

Rules from Belief Networks: A Rough Set Approach

A new version of the Belief SEEKER software that incorporates some aspects of rough set theory is discussed in this paper. The new version is capable of generating certain belief networks (for consistent data) and possible belief networks (for inconsistent data). Then, both types of networks can be readily converted onto respective sets of production rules, which includes both certain and/or possible rules. The new version or broadly speaking-methodology, was tested in mining the melanoma database for the best descriptive attributes of skin illness. It was found, that both types of knowledge representation, can be readily used for classification of melanocytic skin lesions.

Infoscience technology: the impact of internet accessible melanoid data on health issues

In this paper the development of a new internet information system for analyzing and classifying melanocytic dat, is briefly described. This system also has some teaching functions, improves the analysis of datasets based on calculating the values of the TDS (Total Dermatoscopy Score) (Braun-Falco, Stolz, Bilek, Merkle, & Landthaler, 1990; Hippe, Bajcar, Blajdo, Grzymala-Busse, Grzymala-Busse, & Knap, et al., 2003) parameter. Calculations are based on two methods: the classical ABCD formula (Braun-Falco et al., 1990) and the optimized ABCD formula (Alvarez, Bajcar, Brown, Grzymala-Busse, & Hippe, 2003). A third method of classification is devoted to quasi-optimal decision trees (Quinlan, 1993). The developed internet-based tool enables users to make an early, non-invasive diagnosis of melanocytic lesions. This is possible using a built-in set of instructions that animates the diagnosis of the four basic lesions types: benign nevus, blue nevus, suspicious nevus and melanoma malignant. This system is available on the Internet website: http://www.wsiz.rzeszow.pl/ksesi.

A Comparison of Six Discretization Algorithms Used for Prediction of Melanoma

Melanoma is a very serious and lethal skin cancer. In this paper six discretization algorithms, used for preprocessing of melanoma data, were compared using criteria of rule set complexity, total number of errors, and expert’s evaluation. The best discretization method was based on divisive clustering technique. An additional experiment in which the best rules from all six rule sets, selected by an expert, were used for melanoma prediction, was additionally conducted. Our conclusion is that in the original data set, cases with suspicious melanoma were not well represented.

A Search for the Best Data Mining Method to Predict Melanoma

Our main objective was to decrease the error rate of diagnosis of melanoma, a very dangerous skin cancer. Since diagnosticians routinely use the so-called ABCD formula for melanoma prediction, our main concern was to improve the ABCD formula. In our search for the best coefficients of the ABCD formula we used two different discretization methods, agglomerative and divisive, both based on cluster analysis. In our experiments we used the data mining system LERS (Learning from Examples based on Rough Sets). As a result of more than 30,000 experiments, two optimal ABCD formulas were found, one with the use of the agglomerative method, the other one with divisive. These formulas were evaluated using statistical methods. Our final conclusion is that it is more important to use an appropriate discretization method than to modify the ABCD formula. Also, the divisive method of discretization is better than agglomerative. Finally, diagnosis of melanoma without taking into account results of the ABCD formula is much worse, i.e., the error rate is significantly greater, comparing with any form of the ABCD formula.

Optimization of the ABCD Formula for Melanoma Diagnosis Using C4.5, a Data Mining System

Our main objective was to improve the diagnosis of melanoma by optimizing the ABCD formula, used by dermatologists in melanoma identification. In our previous research, an attempt to optimize the ABCD formula using the LEM2 rule induction algorithm was successful. This time we decided to replace LEM2 by C4.5, a tree generating data mining system. The final conclusion is that, most likely, for C4.5 the original ABCD formula is already optimal and no further improvement is possible.