AIDA: uma arquitetura inteligente para gerenciamento de diabetes

Abstract

The incidence of type 2 diabetes mellitus has increased significantly in recent years, emerging as a significant cause of morbidity and mortality. The disease management system currently focuses on identifying a disease and its curable methods. This approach is not suitable for patients with DM2, as the disease is chronic, and the approach must be concerned with the long course of the disease and with everyone involved in its follow-up (the patient, the primary care service, doctors, and hospitals). The evolution of portable devices and sensors and the increased use of electronic health records contribute to the increase in the volume of data that can be used to monitor, improve, and individualize treatment. The extensive data availability for each patient makes the analysis process by health services challenging. This fact encourages the use of artificial intelligence (AI) techniques to extract knowledge that can be used to support healthcare professionals’ decisions. Considering that care for chronic diseases is long-term. It requires continuous monitoring, we propose as a contribution to this work the development of a computational architecture that, based on data collected by sensors and data from the patient’s electronic health record, uses machine learning to find individualized patterns of the course of the disease. The aim is to detect the emergence of comorbidities early and the consequent decline in the patient’s health. For this thesis, we propose developing, implementing, and evaluating an intelligent architecture called AIDA for monitoring patients with type 2 diabetes mellitus. We evaluated the model through technology acceptance assessment studies based on the TAM (technology acceptance model) and the application of the SUS (System Usability scale) in interviews with experts and through the use of technology by patients. Regarding classification and prediction evaluations, we built a dataset with data from a national clinical center. To evaluate the results, several machine learning models for classification were compared, such as Random forest, Decision Tree, Logistic regression, Gradient boosting, XGBoost and LightGBM with hyperparameter settings changed to get the best results. LSTM networks (Long short-term memory) were used for the prediction model with data collected in the patient’s context. The results were evaluated visually and by metrics such as accuracy, precision, sensitivity, specificity, F1-score, and AUROC for classification and Root mean square error (RMSE), Mean absolute error (MAE ) and Mean Poisson Deviance (MPD) for prediction. The acceptance and usability results showed that both groups, professionals and patients, have positive perceptions about applying technology in the care of chronic diseases. In turn, the best results for classification presented AUC of 0.85, specificity of 0.92, sensitivity of 0.52, precision of 0.46, accuracy of 0.87 and F1-score of 0.49. For prediction, the best results presented RMSE of 38.74, MAE of 31.41, and MPD of 8.82. The results reinforce the hypothesis that it is possible to define a computational model to support and monitor patients with type 2 diabetes mellitus.