РАЗРАБОТКА ПРОГНОСТИЧЕСКОЙ МОДЕЛИ ДЛЯ ВЫБОРА МЕТОДА КИШЕЧНОГО АНАСТОМОЗА У ДЕТЕЙ С НАИЛУЧШЕЙ ПРОГНОСТИЧЕСКОЙ СИЛОЙ НА ОСНОВЕ ROC-AUC

Abstract

Objective of the study: development of mathematical models to assess the method of formation of the intestinal anastomosis with a comparison of the effectiveness of these models. Materials and methods of research: from January 2012 to December 2019, 204 operations were performed to restore the continuity of the intestinal lumen, of which 126 were performed using a stapler, 78 – using manual techniques. 90 of the 204 patients were girls (44.1%) and 114 were boys (55.9%). The age of the patients varied from 1 to 322 days, weight – from 950 to 7900 g, the minimum gestational age – from 22 to 42 weeks. To determine the method of intestinal anastomosis formation, the following mathematical models were used: binary logistic regression, binary logistic regression with the method of stepwise inclusion based on maximum likelihood (Forward: LR), neural network. In the first method of calculations, all variables declared as covariates were simultaneously involved in the calculation. The second method – Forward: LR, started by using only constants at the start, and then sequentially connected variables that demonstrate a strong relationship with dependent variables. The third method is a neural network. A radial basis function was used as a mathematical basis for constructing a neural network. To train the neural network, the active dataset was divided into training, validation and control samples. Observations were randomly divided based on relative numbers (50%, 25% and 25%). Results: 3 patients with different initial data were selected for a comparative analysis of the three mathematical models. Evaluation of the effectiveness of the obtained mathematical models was based on the construction of ROC curves. It was found that for the binary logistic regression the area under the ROC curve was 0.975, for the binary logistic regression with the stepwise inclusion method based on the maximum likelihood (Forward: LR) the area under the ROC curve was 0.975, for the radially basic neural network the area under the ROC curve amounted to 0.888. Conclusions: the quality of the doctor's work can be confirmed by methods of mathematical modeling. The choice of surgical intervention is implemented by three mathematical models and their qualitative assessments are given. The developed integrated approaches can be made available to specialists in subject areas for use in clinical practice, as well as for machine learning and the creation of future automated systems based on intelligent mathematical technologies.