Deterioração Clínica em Unidades de Terapia Intensiva: uma Abordagem de Predição de Situação dos Pacientes Baseada em Inteligência Computacional

Abstract

Patients in hospital Intensive Care Units (ICUs) require constant monitoring of their physiological parameters due to the severity of their health condition, coupled with the possibility of rapid deterioration. Given this motivation, this thesis aims to conceive an approach for predicting clinical deterioration of ICU patients called I2Care. Thus, the research developed focuses on the balance between the classification per formance obtained through Machine Learning models and the interpretability offered by Fuzzy rules. To evaluate the work using real patient data, the MIMIC-III database was selected, considering its broad recognition within the community. Among Machine Learning-based predictive models, the best performance was achieved by I2Care-ML, which explores the Gradient Boosting Machine algorithm, reaching an AUC of 0.8846 using ten variables. On the other hand, among Fuzzy Logic-based predictive models, I2Care-DS performed the best. Using only five clinical variables, it achieved a significant performance, reaching an AUC of 0.8651. This result not only demonstrates better predictive performance for the conceived approach but also highlights its superiority compared to studies in the literature using Fuzzy Logic. Furthermore, the I2Care approach showed superior predictive performance compared to SAPS-II, a widely used clinical score in hospitals. These results suggest that I2Care not only surpasses other methods in the literature in terms of predictive performance but also stands out for reducing the use of predictor variables and the ability to infer the degree of clinical deterioration in patients. These properties of the conceived approach can have practical implications in terms of reducing data collection, lower implementation costs, as well as better clinical interpretability of patients’ situations.