Predição de mortalidade em UTI: aplicação de técnicas de mineração de dados

Abstract

The use of standardized scores to identify the severity of the condition of patients admitted to the Intensive Care Unit - ICU, such as Acute Physiology, Age, Chronic Health Evaluation APACHE III and Simplified Acute Physiology Score - provides information used by the medical team to make decisions, these severity scores go through constant revisions that seek to improve their predictive capacity, due to using linear methodologies for prediction and the data used to obtain the scores have non-linear characteristics, we understand that it can be used other techniques and methodologies to improve the prediction of these scores. This study aims to propose the application of data mining methods, in the preprocessing of the entire database and in the identification of the severity of the patient’s condition using Neural Networks RNA, Random Forest - RF and Logistic Regression, having as attributes to analyze the records of the physiological variables already registered by the medical team to calculate the mentioned scores. The data used for this purpose were obtained from the Medical Information Mart for Intensive Care (MIMIC-III), a large available on-line repository for searches containing a record of 56,530 patients. In addition, we analyzed techniques of imputation of missing values and class balancing, in the search for a higher quality in the data. After application of the methodology described in the study Random Forest obtained better performance than the others, with mean AUC of 0.780 (± 0.005), sensitivity of 0.712 (± 0.012) and specificity of 0.701 (± 0.005) in conjunction with the technique of imputation of standard values in substitution of missing values, and with class balancing using under sampling. By means of attribute selection, a model with attributes reduction was created with results close to the classification with all attributes, which can simplify the data collection by the medical team to generate a severity score.