Machine learning models in decision support systems for diagnosing colorectal cancer based on metabolic profiles

• Autor(a) principal: Barbosa, Rui Xavier Ferreira
• 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/10400.22/24307
• topic: Colorectal Cancer; Machine Learning; Amino Acids; Acylcarnitines; ExplainableAI; Domínio/Área Científica::Engenharia e Tecnologia

In today’s ever-evolving technological landscape, the volume of data across sectors is grow ing, particularly in healthcare. Here, the gathering and processing of biochemical data aim to refine decision-making for patient treatments, especially using tools based on Machine Learning (ML). As a subset of Artificial Intelligence, ML harnesses algorithms to predict outcomes or unearth patterns that might otherwise remain concealed. The interpretability of ML models is pivotal, enabling healthcare professionals to place con fidence in and decipher the model’s predictions. This assumes particular significance when decisions could directly affect patient lives. This research embarked on an in-depth exploration of various ML algorithms and techniques to discern whether the combined metabolic profiles of amino acids and acylcarnitines might serve as new biochemical indicators for predicting colo-rectal cancer prognosis. Throughout this study, several algorithms and data preprocessing techniques were evaluated. Four distinct experiments validated the predictions of the models in different scenarios. These scenarios involved predicting Colorectal Cancer using amino acids with and without the age parameter, and similarly, using acylcarnitine with and without the age parameter. Each scenario’s predictions were elucidated using SHAP, both for overarching feature significance and individual instances. Preliminary analyses indicated that the constructed models demonstrated promising predic tive power, with notable variations for the different scenarios. Amongst the algorithms tested, Random Forest, Support Vector Machine, Gaussian Naive Bayes, and Gradient Boosting emerged as the top performers.