Machine Learning for the Early Detection of Acute Episodes in Intensive Care Units

• Autor(a) principal: Ribeiro, Bruno Marques
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/143720
• topic: Intensive Care; Physiological Signals; Acute Hypotensive Episodes; Anomaly Detection; Machine Learning; Layered Learning; Domínio/Área Científica::Engenharia e Tecnologia::Outras Engenharias e Tecnologias

In Intensive Care Units (ICUs), mere seconds might define whether a patient lives or dies. Predictive models capable of detecting acute events in advance may allow for anticipated interventions, which could mitigate the consequences of those events and promote a greater number of lives saved. Several predictive models developed for this purpose have failed to meet the high requirements of ICUs. This might be due to the complexity of anomaly prediction tasks, and the inefficient utilization of ICU data. Moreover, some essential intensive care demands, such as continuous monitoring, are often not considered when developing these solutions, making them unfit to real contexts. This work approaches two topics within the mentioned problem: the relevance of ICU data used to predict acute episodes and the benefits of applying Layered Learning (LL) techniques to counter the complexity of these tasks. The first topic was undertaken through a study on the relevance of information retrieved from physiological signals and clinical data for the early detection of Acute Hypotensive Episodes (AHE) in ICUs. Then, the potentialities of LL were accessed through an in-depth analysis of the applicability of a recently proposed approach on the same topic. Furthermore, different optimization strategies enabled by LL configurations were proposed, including a new approach aimed at false alarm reduction. The results regarding data relevance might contribute to a shift in paradigm in terms of information retrieved for AHE prediction. It was found that most of the information commonly used in the literature might be wrongly perceived as valuable, since only three features related to blood pressure measures presented actual distinctive traits. On another note, the different LL-based strategies developed confirm the versatile possibilities offered by this paradigm. Although these methodologies did not promote significant performance improvements in this specific context, they can be further explored and adapted to other domains.