Análise exploratória sobre registros eletrônicos de saúde do setor de unidade de terapia intensiva utilizando modelagem de tópicos

Abstract

The rapid growth of electronic health record systems brings the increase of available information about patients in hospitals. This massive amount of text information is suitable for the extraction of unknown information about medical history, medication, deseases, allergies, among others. Topic modeling is a machine learning problem, which aims to extract, given a collection of documents, the main topics that represent the subjects covered by a text collection. In the topic model, the documents can be composed of a mixture of topics with a certain probability. This work aims to make an exploratory analysis of two collections of electronic records health from an intensive care unit. The collection is split into two subcollections: discharged patients and patients who progressed to death. We apply the Latent Dirichlet Allocation (LDA) algorithm in both collections, setting the number of topics to 11. As a result, discharged patients collection shows the following predominant topics: respiratory system, renal system, neurological system, prematurity, and cardiac system. On the other hand, the death collection presents as main topics subjects about the hepatic system, cardiovascular, neurological, and respiratory system. We also analyze the correlation of the topics inter collections, and we observed that the infection as a significant contributing factor to progress to death.