Inteligência artificial aplicada ao ecossistema de regulação do Estado Rio Grande do Norte (RegulaRN): análises baseadas em machine learning em leitos Covid-19 e leitos gerais

Abstract

The process of bed regulations is among the most relevant processes for the Brazilian public health system. It encompasses the entire process of managing and monitoring a patient who requires hospitalization, from the request to their proper admission. However, it is still an area that has little investment in digital health systems and other resources that can favor the better management of the regulatory process. Thus, this work aims to include the area of artificial intelligence within the area of regulating public beds, in order to enhance and assist the decision-making process during bed regulation. In this sense, bed regulation data from two modules of the platform adopted in Rio Grande do Norte, RegulaRN COVID-19 and RegulaRN Leitos Gerais, were used in order to classify the data and predict the patient's outcome. In total, approximately 72,422 bed regulation data were analyzed in different time frames. In addition, a pipeline of characterization, preprocessing, data correlation, definition of metrics for evaluation, data balancing, definition of training and validation data, definition of computational models for data classification and selection of hyperparameters was used. For the RegulaRN COVID-19 platform, the results showed better performance for the accuracy (84.01%), precision (79.57%) and F1-score (81.00%) metrics in the Multilayer Perceptron (MLP) model with Stochastic Gradient Descent (SGD) optimizer. For the recall (84.67%), specificity (84.67%) and ROC-AUC (91.6%) metrics, the best results were obtained by Root Mean Squared Propagation (RMSProp). As for the RegulaRN Leitos Gerais data, the analyses were performed with two datasets: adults and pediatric and neonatal. For the first set, Extreme Gradient Boosting (XGBoost) presented the best accuracy (87.77%) and recall (87.77%) values, Random Forest had the best precision (87.05%), Gradient Boosting had the best F1 Score (87.56%) and for specificity (82.94%) it was obtained by SGD. For the newborn data, the best accuracy (87.50%), recall (87.50%) and F1-Score (88.48%) values were obtained by the Decision Tree classifier, the best precision (90.75%) by Adaboost and the best specificity by MLP Adam. The results allowed us to identify the best models to assist health professionals during the bed regulation process, as well as the scientific findings of this academic work demonstrate that the computational methods used applied through a digital health solution can assist in the decision-making of medical regulators and government institutions in order to strengthen the performance of Brazilian public health.