UNCERTAINTY IN MACHINE LEARNING A SAFETY PERSPECTIVE ON BIOMEDICAL APPLICATIONS

• Autor(a) principal: Barandas, Marília da Silveira Gouveia
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/162131
• topic: Machine learning; Uncertainty quantification; Classification with rejection option; Interpretability; Clinical decision making; Domínio/Área Científica::Engenharia e Tecnologia::Outras Engenharias e Tecnologias

Uncertainty is an inevitable and essential aspect of the worldwe live in and a fundamental aspect of human decision-making. It is no different in the realm of machine learning. Just as humans seek out additional information and perspectives when faced with uncertainty, machine learning models must also be able to account for and quantify the uncertainty in their predictions. However, the uncertainty quantification in machine learning models is often neglected. By acknowledging and incorporating uncertainty quantification into machine learning models, we can build more reliable and trustworthy systems that are better equipped to handle the complexity of the world and support clinical decisionmaking. This thesis addresses the broad issue of uncertainty quantification in machine learning, covering the development and adaptation of uncertainty quantification methods, their integration in the machine learning development pipeline, and their practical application in clinical decision-making. Original contributions include the development of methods to support practitioners in developing more robust and interpretable models, which account for different sources of uncertainty across the core components of the machine learning pipeline, encompassing data, the machine learning model, and its outputs. Moreover, these machine learning models are designed with abstaining capabilities, enabling them to accept or reject predictions based on the level of uncertainty present. This emphasizes the importance of using classification with rejection option in clinical decision support systems. The effectiveness of the proposed methods was evaluated across databases with physiological signals from medical diagnosis and human activity recognition. The results support that uncertainty quantification was important for more reliable and robust model predictions. By addressing these topics, this thesis aims to improve the reliability and trustworthiness of machine learning models and contribute to fostering the adoption of machineassisted clinical decision-making. The ultimate goal is to enhance the trust and accuracy of models’ predictions and increase transparency and interpretability, ultimately leading to better decision-making across a range of applications.