Bayesian Inference in stochastic process to identify mortality attributed to sepsis

• Autor(a) principal: Eugenio, Nicholas Wagner
• Data de Publicação: 2024
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: https://www.teses.usp.br/teses/disponiveis/45/45133/tde-12082024-083827/
• topic: Bayesian; Bayesiana; Cadeia de Markov não homogênea; Fração populacional atribuível; Non-homogeneous Markov chain; Population attributable fraction; Processo estocástico; Stochastic process

This work introduces a new approach to calculating attributable population fractions (PAF) and attributable hazard functions (AHF) within the context of stochastic processes and non-homogeneous Markov chains, aiming to reconcile this approach with existing literature. It begins by discussing the concept, motivation, and origin of PAF, highlighting its flexibility in different study designs in Chapter 1. A national-scale study, involving over 3800 hospitalized patients across 38 medical centers, and relating exposure to sepsis over a period of hospitalization to the outcomes of death and discharge, served as the motivation for developing the new presented approach, as exposed in Chapter 2. Chapter 3 provides a comprehensive overview of calculating PAF and AHF, including time-dependent variations. Chapter 4 delves into the Bayesian estimation of transition probabilities in Markov chains, covering both homogeneous and heterogeneous cases. Our proposed Adapted Attributable Hazard Fraction (AAHF) is introduced in Chapter 5, incorporating failure rate formulas from competing risks analysis, the theory of non-homogeneous discrete Markov chains, and survival analysis. This allows for the calculation of metrics for both a specific set of covariates (subpopulation) and general measures. The new approach enables inference on the transition probabilities between the observed states of individuals and applies these in constructing competing risk failure rates, which, in turn, structure the AAHF formula for the outcome under study, in this case, patient mortality. Moreover, it is innovative in being a dynamic measure over time, as well as the effects of covariates. Chapter 6 applies this new approach to a filtered dataset from the motivational study, observing transitions in patient outcomes (discharge and death) and risk factors (sepsis and non sepsis) over time. The results of the average AAHF are presented. Finally, in Chapter 7 we have the concluding remarks of the study that sheds light on the delayed impact of sepsis on mortality in hospitalized patients. Initially (days 1-13), there is no significant difference in mortality attributable to sepsis exposure, suggesting effective early interventions or unidentified high-risk patients. From day 14 to 17, sepsis-related mortality emerges, with 1% attributed to sepsis exposure, indicating worsening effects with prolonged hospitalization. From day 18 onwards, mortality attributed to sepsis exposure rises to approximately 2%, emphasizing the need for continuous monitoring and aggressive sepsis management in long-term hospitalized patients.