Predição em análise de sobrevivência: aplicação em estudo envolvendo óbito de pacientes chagásicos cardiopatas

• Ano de defesa: 2019
• Autor(a) principal: Emilly Malveira de Lima
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ICX - DEPARTAMENTO DE ESTATÍSTICA; Programa de Pós-Graduação em Estatística; UFMG
• Palavras-chave em Português: dadis longitudinais e de sobrevivência; modelo de Cox; modelagem conjunta; landmark; predição de risco; Estatística – Teses; Teoria da previsão – Medicina - Teses; Análise de sobrevivência (Biometria) - Teses; Doença de Chagas - Teses.
• Link de acesso: http://hdl.handle.net/1843/51001

Risk prediction models have been used in many areas, especially in the medical field. In many situations, they can be used as a clinical aid tool to help, for example, to define severity risk groups, and decide about the most appropriate treatment. There are circumstances where the interest may be, for example, to predict the risk of a patient being death at a given time, incorporating variables (clinical markers) recorded at baseline and over time. This work was motivated by the need to construct a risk score for patients with Chagas cardiomyopathy from the prospective SaMi-Trop cohort. Patients residing in 21 municipalities of the northern Minas Gerais were followed, initially, for two years and subsequently, it was of interest to create a two-year death risk score with baseline information. The follow-up of the study allowed the nature of the risk of death to be dynamic, whereas there are clinical markers that change over time, and therefore the risk score needs to be updated. In this article, we investigate the use of four approaches: Naive 1, Naive 2, landmark (LM ), and joint modelling (JM ). These methods are based on the Cox regression model and the class of models for longitudinal and survival data. It was performed static and dynamic predictions, and the performance of each method was assessed using discrimination and calibration measures. The suggested results for all approaches have satisfactory discrimination, however, not all shown have a good calibration and returned inaccurate probabilities.