Генетические алгоритмы определения высокоинформативных признаков заболеваний молочных желез

Genetic algorithms for determination of the highly informative signs of mammary glands deseases

A.G. Losev, E.A. Mazepa and T.V. Zamechnik in a recent paper [5] proposed an algorithm for obtaining highly informative primary diagnostic features of breast health based on microwave radiometry. The primary diagnostic features are based on the analysis of numerical functions describing the known high-quality medical signs identified by experts-mammologists in research data obtained during breast examination (see. [8]). For example, the increased value of thermo-asymmetry between the same points of the breast can be described by functions of the form |􀝐пр.􀯜-􀝐л.􀯜|, where 􀝐пр.􀯜 and 􀝐л.􀯜 are temperatures at i points on right and left breasts respectively. To characterize the quality of the diagnostic feature A.G. Losev introduced the concept of the combined informativeness. The higher value of combined informativeness of classification is the better sign for defining the difference between required and separated groups. This article explores the possibility of obtaining more informative signs based on linear combinations of previously obtained primary symptoms. Selecting weights by the genetic algorithm in the mentioned combinations, we can obtain signs with twice-higher informativeness than in the primary ones. Initial symptoms can be divided into groups, each of which describes a qualitative clinical symptom. We investigated the linear combinations of primary features that describe the following qualitative medical symptoms: reduced value of nipple temperature compared with the temperatures of neighboring points, a reduced value of the difference between deep and surface temperatures at several points of the breast, increased dispersion of surface temperatures between separate points in the affected mammary gland, increased dispersion of deep temperatures between separate points. In each group, we found new features with greater informativeness than the primary characteristics of the group. We attempted to combine the primary signs of the distant groups in terms of medicine. In these combinations we obtain the largest combined informativeness. However, these combinations of symptoms have no medical justification. It is possible that these symptoms just track the nuances of training sample.