Modelo preditivo do nível glicêmico por monitoramento em tempo real em indivíduos portadores de diabetes mellitus tipo II

Abstract

This research deals with a prediction of the glycemic levels of people with Diabetes Mellitus II, collected through a continuous glycemic monitoring system, based on the architecture of LSTM neural networks. Diabetes, one of the non-communicable chronic diseases, is characterized by hyperglycemia in the bloodstream generated by insulin resistance. The control of this disease can occur through carbohydrate counting according to the glycemic level, which according to the anthropometric evaluation is quantified by the physician. However, this approach is not always well accepted by diabetics, who end up adhering to medication for their control. Despite this, some diabetics end up using continuous blood glucose monitoring sensors, which favored verifying whether the glycemic data collected every 15 minutes could be predicted. The glycemia of 20 patients was measured over a period of 14 days using real-time monitoring. During this period, eating habits were recorded to count ingested carbohydrates, using the carbohydrate counting app created by SBD. Using an artificial intelligence model (LSTM) a blood glucose prediction model was created. With this model, it was verified that the predicted values followed the real glycemic movement, anticipating 5 hours with glycemic data of 12 continuous hours, that is, 20 predicted observations and 48 observations collected by the glycemia sensor for each individual. A general predictive model was performed with 20 volunteers and two personalized ones. The glycemic data of the collected diabetics had a positive performance, as the predicted values followed the glycemic movement, with a glycemic peak of 170 mg/dL at 9 am and 180 mg/dL at 1 pm, converging with the data obtained from the blood count. of carbohydrates, physical and anthropometric evaluation, observed with the peaks of glycemia, lifestyles of the volunteers and the total carbohydrates consumed daily. The glycemic data of non-diabetics had a positive performance, given that the predicted data followed the actual glycemic movement. This model, therefore, can predict several applications directly in rehabilitation, contributing as one of the important instruments for improving the patient's quality of life.