**4. Discussion**

In the previous section, we described the application of the proposed approach to solving practical medical classification problems.

The first dataset describes tissue characteristics to diagnose benign or malignant breast neoplasm. This dataset was studied many times before, for example [69,70], including the approach based on LAD [2]. For this example, we established a significant influence of the introduced hyperparameter of the genetic algorithm—the probability of dropping one during the initialization of the initial population, which is the fixation of the attribute's value in the pattern according to its value in the baseline observation. High values of this probability lead to low classification accuracy due to excessive selectivity of patterns and, accordingly, an increase in the number of objects with a refusal to determine the class membership.

The second dataset is the problem of predicting complications of myocardial infarctionatrial fibrillation. In this case, two approaches are implemented to handle missing values in the data. The first, typical for most data mining algorithms, is a combination of deleting values with missing values and filling them in. In this case, satisfactory classification results were not achieved when a complete dataset with a significantly larger presence of objects of one of the classes was used. In the second approach, the missing values are not preprocessed since the set of logical patterns as a whole has no restrictions in classifying observations with missing values. In addition, we use a reduced sample with an equal number of observations for each of the classes.

Homogeneous patterns in this dataset have small coverage, and using only homogeneous patterns leads to overfitting. Relaxation of homogeneity constraints requires adjusting the threshold (right-hand side of the constraint), which can be difficult since, when solving pattern finding problems, the best balance between coverage and homogeneity for a single pattern can be far from the best for another pattern (based on another baseline observation). The proposed approach simplifies the search for this balance since it considers many Pareto-optimal patterns. This approach prevents overfitting in contrast with using only homogeneous patterns or patterns with a given threshold for homogeneity. At the same time, the accuracy reaches the values obtained using an artificial neural network specially developed for these data [67]. The classification results are also comparable with the results of other works on this topic [71,72].
