Previsão de mortalidade após cirurgia cardíaca congênita utilizando aprendizagem de máquina

Abstract

Congenital heart disease is the most common cause of major anomalies of the same gender, accounting for almost a third of all major congenital anomalies. Congenital heart defects are serious and common conditions with a significant impact on morbidity, martality and health costs for children and adults. In the treatment of patients with congenital heart disease, research related to the risk of pre-surgical mortality is rare. This study aims to propose a model ops individual risk of death prediction for cardiac surgery of patients with congenital heart disease and to assist health professionals in understanding which diagnoses or variables are assoaciated with the risk of death. Teh use of machine learning techniques as a tool to suppoort decision making in the field of medicine has been increasing in recent years. With the information on surgeries performed on patients with congenital heart disease extracted from the ASSIST database of InCor, it was possible to rtain six different machine learning algorithms in predictiong the risk of pre-surgical mortality and to understand which variables impact the risk death of these patients. The algorithms trained inthis study were: Miltilayer Perceptron (MLP), Random Forest (RF), Extra Trees (ET), Stochastic Gradient Boosting(SGB), AdaBoost Classification (ABC) and Bagged Decision Trees (BDT). To predict the risk of patient mortality, the model with the best performance was the Random Forest (RF) with ROC AUC (area under the receiver's operating characteritics) of 90,2%, AP indexes (average precision) 0f 0,73 and sensitivity index (recall) mof 92,2%. The machine learning algorithm (machine learning0 can assist in understanding the mortality risks of patients with congenital heart disease when undergoing cardiac surgery and using clinical drugs that understand the best risks associated with surgical interventions, providing information to support the decision, health professionals, patients and their families