Diagnostic Options for Postoperative Complications in Surgery

Abstract

Currently, scientific research is usually carried out in accordance with the postulates of evidence-based medicine in Russia and foreign countries. However, the implementation of these principles requires deep knowledge of surgery and mathematical modeling. Authors: a surgeon and a programmer developed mathematical models involved in the diagnosis of postoperative complications in surgery. In this paper, we investigated a deep, fully connected neural network for the diagnosis of postoperative complications on the clinical example of acute appendicitis. As a training set of parameters, we used a set developed by the authors on the basis of real clinical data, which has a state registration number in the form of a database, and includes a knowledge base. The interquantile range of the F1 measure is proposed for the selection of significant features. An approach to coding composite categorical features, characterized by a compact representation, is proposed. For pre-processing of training data, it is proposed to use a step-up autoencoder. The autoencoder converts the selected functions into a higher-dimensional space, which, according to Kover's theorem, facilitates the classification of features. The neural network is implemented using the Keras and TensorFlow libraries. To train the neural network, the Adam algorithm with adaptive learning speed is used. To reduce the effect of overfitting, a modern regularization method dropout-was used. The analysis and selection of the classifier quality metrics are carried out. To evaluate the characteristics of the neural network, k-block cross-validation was used. The trained neural network showed high diagnostic performance on the test data set.