Estratégias de machine learning para predição de mortalidade em traumatismo cranioencefálico: uma revisão sistemática da literatura

• Ano de defesa: 2023
• Autor(a) principal: Arthur Afonso Silva-Sousa
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ICB - DEPARTAMENTO DE MORFOLOGIA; Programa de Pós-Graduação em Neurociências; UFMG
• Palavras-chave em Português: Traumatismo cranioencefálico; Mortalidade; Machine learning; Revisão sistemática; Neurociências; Traumatismos Craniocerebrais; Aprendizado de Máquina; Revisão Sistemática
• Link de acesso: http://hdl.handle.net/1843/57467

Traumatic brain injury (TBI) has been recognized as one of the main causes of morbidity and mortality worldwide, being considered a silent epidemic. A variety of causes of TBI are considered, such as falls, car accidents, robberies and assaults. Despite intense investments, there is not still an approved treatment to mitigate the damage caused by secondary TBI injuries, and prevention is currently the best treatment. With technological advances, machine learning (ML) techniques have been used in the health field to predict outcomes and aid in therapeutic decision-making, for example, mortality risk in TBI victims. However, to the same extent that there is a significant development of ML techniques for this purpose, there is no consensus on which strategies have the best performance in predicting mortality. Based on the above, the objective of this study was to carry out a systematic review of scientific productions that use machine learning techniques to predict mortality due to TBI. To this end, the research question was formulated from the PICO strategy (Person, Intervention, Comparison, Outcome), having as inclusion criteria scientific articles published in journals indexed in the Virtual Health Library (BVS) and the National Library of Medicine (PubMed), written in English, Portuguese and Spanish, that had mortality as one of the study outcomes. Studies with animal models, focusing on molecular analyzes and articles that were not found for reading in full were excluded. Data collection took place in April 2022, by two independent researchers and, for the analyses, a specific protocol was constructed. Of the 1181 studies found, 26 met the eligibility criteria and were included for analysis in this review. The largest number of publications occurred among the Four Asian Tiger countries (n=7, 26.9%), with retrospective studies (n=16, 61.5%) and adult sample (n=19, 73.1% ). The overall mean age was of 47.8, with a predominance of males among TBI victims and the main causes were traffic accidents, falls and objects hitting the head. The main variables that predicted mortality were age and pupillary response (n=10, 9.4%) and low scores on the Glasgow Coma Scale (n=9, 8.5%). Fifty-nine machine learning algorithms were found, with the best performance being Artificial Neural Networks, Support Vector Machine, and Random Forest, especially when compared to the traditional logistic regression algorithm. Based on the results obtained, it can be preliminarily concluded that LM strategies are a useful and effective tool for predicting mortality in TBI, favoring clinical decision-making. As a perspective of this study, it is proposed to assess the risk of bias, update the search, including the artificial intelligence descriptor and perform a meta-analysis, as the data allow.