Нейросетевое прогнозирование гликемии у пациентов с сахарным диабетом на основе смешанных временных рядов с перспективой применения в составе интеллектуальной инсулиновой помпы

Abstract

The paper considers the prospect of using a neural network self-learning algorithm for personalizing insulin therapy within the software and hardware complex of an intelligent insulin pump (IIP). IIP is an autonomous automatic insulin control system that can completely simulate the secretion of insulin by the pancreas of a healthy person. The required dose calculation is carried out on the basis of biological data from sensors installed on the patient's body. The algorithm for calculating the dose is an intelligent self-learning neural network unit that allows you to achieve fine personalization of therapy with the possibility of flexible readjustment in accordance with the dynamically changing biological parameters of the patient in real time. The article discusses several possible neural network paradigms for predicting blood sugar in the short-term (3 minutes) and medium-term (30 minutes) periods based on a mixed time series that includes measurements of blood glucose, active insulin and active carbohydrates at intervals of 3 minutes. The advantage of using MLP (multilayer perceptron) networks over other paradigms, in particular, LSTM networks (long short-term memory networks), is shown. The results of computational experiments using a neural network model on real data of two volunteers with different degrees of insulin sensitivity are presented. The impossibility of unifying the model for patients with different sensitivities has been proven, which confirms the need to personalize therapy for insulin-dependent diabetes. On the basis of the results of the experiments, recommendations are given on the construction and training of neural network models for predicting blood glucose levels in patients, as well as prospects and directions for further research.