Modelos preditivos para sepse: um estudo de reprodutibilidade, padronização e confiabilidade em aprendizado de máquina

Abstract

Sepsis is a serious medical condition characterized by the body’s systemic in flammatory response to infection and is associated with a high mortality rate, making early detection essential for successful treatment. In recent decades, machine learning models have been widely explored to predict its occurrence, yet the literature still suffers from major issues related to reproducibility, comparability, and the lack of methodological standardization. This dissertation proposes a critical evaluation of the state of the art in sepsis prediction, alongside the development of a stan dardized framework for testing and comparing models. To this end, a standardized dataset was created based on MIMIC-IV v2.2, using the Sepsis-3 definition, with reproducible clinical criteria and well-defined time windows. State-of-the-art models were reimplemented and evaluated both in their original contexts and within this controlled environment. Additionally, this research incorporated the Conformal Predic tion technique, in its transductive form, as a means of quantifying uncertainties and enhancing the reliability of predictions. The results showed that reproducibility remains a challenge, that model performance varies significantly across different time windows, and that the application of conformal prediction can improve the safety and precision of predictions, although its effectiveness depends on calibration and the nature of the underlying model. The contributions of this dissertation include the standardization of experimental practices, a critical analysis of the current literature, the demonstration of limitations in existing models, and the proposal of more robust and reliable pathways for the use of AI in sensitive clinical settings.