CASPIAN JOURNAL
MANAGEMENT AND HIGH TECHNOLOGIES
THE ALGORITHM FOR ASSESSING THE PERFORMANCE OF NEURAL NETWORKS IN SOLVING CLASSIFICATION PROBLEMS IN MEDICAL DIAGNOSIS
 |
Read | Sidorova Margarita A., Serzhantova Natalia A. THE ALGORITHM FOR ASSESSING THE PERFORMANCE OF NEURAL NETWORKS IN SOLVING CLASSIFICATION PROBLEMS IN MEDICAL DIAGNOSIS // Caspian journal : management and high technologies. — 2017. — №4. — pp. 9-19. |
/9-19.jpg)
Sidorova Margarita A. - Cand. Sci (Engineering), Associate Professor, Penza State Technological University, Baydukov Av. / 1a/11 Gagarin St., Penza, 440039, Russian Federation, sidorova_mailbox@mail.ru
Serzhantova Natalia A. - Cand. Sci (Engineering), Associate Professor, Penza State Technological University, Baydukov Av. / 1a/11 Gagarin St., Penza, 440039, Russian Federation, itmmbspgta@yandex.ru
The article discusses some of the existing criteria for assessing the performance of neural networks (ANNs) to solve classification problems. These criteria include the following: training error functions (sum of squared errors, root-mean-square error, variable error and mean absolute training error); the error of vessel thickness estimation; the probability of first and second type errors occurrence; the share of incorrectly classified objects in the testing sample. The authors reveal the main disadvantages of these criteria. To assess the results of the work of the ANN (exemplified by the problems of classification parameters of hemostasis, such as coagulation and blood viscosity), the paper uses advanced statistical analysis based on Bayesian approach. It goes on to propose an original criterion for assessing the quality of ANN performance to solve classification problems in medical diagnosis; the criterion takes into consideration the threshold value of ANN making prognostically right decision. The authors developed an algorithm that estimates the quality of ANN performance based on the proposed quality criteria. This algorithm allows varying the value of significance level for different tasks, thereby establishing the necessary and sufficient values that define the performance of ANN under study as efficient.
Key words: нейронная сеть, классификация, медицинская диагностика, критерий, качество, уровень значимости, порог принятия решения, эффективность, информативность, ошибка, пропуск, ложное срабатывание, neural network, classification, medical diagnosis, criteria, qual