DEVELOPMENT OF A SOFTWARE COMPLEX FOR CORRECTING THE WEIGHT OF DIABETES PATIENTS BASED ON THE USE OF A NEURAL NETWORK WITH LOGISTIC REGRESSION

Abstract

As a result of the analysis of the field of diabetes prediction, the main factors are identified: heredity, lifestyle, weight, environmental factors, age. The existing methods in diabetes prediction are highlighted: data mining, logistic regression analysis, neural network. The main functionality in the considered applications is chosen. Subject area glossary developed. Mathematical models of neural network activation functions are selected: relu, softmax, sigmoid, linear. The mathematical model for calculating the weights of the neural network is logistic regression, which increases the accuracy of forecasting. A model of the business process of using an artificial neural network to adjust the weight of patients with diabetes, resulting in a technical task for application. Application has been developed to adjust the weight of patients with diabetes using a neural network, which allows you to timely warn the user about high blood sugar, monitor the patient's condition, provide recommendations in the form of a menu for the day, perform analysis of indicators. The use of design templates "Observer" to monitor the doctor's condition of the user, and "Builder" to create a medical card of the user. The work of neural networks "Tensorflow.js" "Brain.js", which are able to provide satisfactory quality of blood sugar and weight forecast, has been studied. Four activation functions (relu, softmax, sigmoid, linear) during training and test predictions of neural networks were studied. It was found that for neural network on the basis of "Tensorflow.js "it is better to apply "relu "to the input neuron, and" softmax "to the output and logistic regression, their combination gives a prediction accuracy of 94%. It was found that for neural network based on "Brain.js" it is better to apply "Linear" activation function, and the accuracy of the forecast is 93%. The average value of the forecast error in the study of neural network "Tensorflow.js" did not exceed - 0.006, and in "Brain.js" - 0.00693. The effectiveness of the use of trained neural networks of direct propagation and recurrent networks to predict blood sugar and weight values is tested. Developed a hybrid model paired with logistic regression, which allowed to achieve accurate prediction of 94% blood sugar for the next 3 - 9 hours, and also scales for 3 - 7 days in advance at the expense of "Tensorflow.js". In most cases, the neural network predicts a sugar level of 3 hours, which is sufficient for a diabetic to take steps to prevent the sugar level from falling or rising.