A hybrid model for long-term prediction of glycemic oscillation in individuals with type 1 diabetes and suggesting personalized recommendations.

• Ano de defesa: 2022
• Autor(a) principal: Pereira, João Paulo Aragão
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Agentes inteligentes; Glucose-insulin regulatory system; Machinel; Multi-agent system; Ordinaryd equations; Personalization; Prediction; Sistemas de recomendação; Sistemas híbridos; Type 1 diabetes
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/3/3141/tde-14032023-092608/

The glucose-insulin regulatory system and its glycemic oscillations is a recurring theme in the literature due to its impact on human life, especially those affected by diabetes mellitus. Several approaches have been proposed, from mathematical models to data-based ones, in order to model the glucose oscillation curve. With this curve, it is possible to predict when and how much to inject insulin, the ideal amount of carbohydrates and possible hyper- or hypoglycemic states in individuals with type 1 diabetes (T1D). However, the literature presents prediction horizons not exceeding six hours, which can be a problem considering the sleeping time. Also, existing models cannot be customized for each individual considering their lifestyle. This work presents Tesseratus, a model that adopts a multi-agent system to combine machine learning and mathematical modeling to predict glucose oscillation in up to eight hours. Tesseratus also uses the pharmacokinetics of insulins, in addition to data collected from individuals with DM1. Periodic data capture can improve the learning process of agents. Its result is essentially glucose prediction values over horizons ranging from 15 to 480 minutes. Therefore, it can assist endocrinologists in prescribing daily treatments for individuals with T1D and providing personalized recommendations for these individuals in order to maintain their blood glucose concentration in the optimal range. Tesseratus brings pioneering results for prediction horizons of eight hours for the night period, in an experiment with 15 real individuals with DM1 and 9 virtual ones. Using the Parkes Error Grid as an evaluation metric, it can be observed that 95.53% of measurements, on average, fall into zones A and B, during the daytime period, while at night it reached 95.1%, with the Mean Absolute Error equal to 26.75 and 27.16 mg/dL, respectively. It is our assertion that Tesseratus will be a reference for the classification of the glycemic prediction model, supporting the mitigation of short and long-term complications in individuals with T1D. In this way, the proposed predictive model tends to delay the acute and chronic complications of a population with a projection of 78 million adults with type 1 diabetes, worldwide, in 2045.