Sistema de suporte à decisão baseado em inteligência artificial para predição de doenças arteriais coronárias

• Ano de defesa: 2022
• Autor(a) principal: Carlos Anderson Oliveira Silva
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ICX - DEPARTAMENTO DE QUÍMICA; Programa de Pós-Graduação em Inovação Tecnológica e Biofarmacêutica; UFMG
• Palavras-chave em Português: Aprendizado de máquina; Doenças coronárias; Sistema de apoio à decisão clínica; Aprendizado fracamente supervisionado; Aprendizado de máquina explicável; Machine learning; Coronary diseases; Clinical decision support system; Explainable machine learning; Weak supervision; Inovações tecnológicas; Inteligência artificial; Aprendizado do computador; Coronariopatias; Sistemas de suporte de decisão; Clínica médica – Processo decisório; Redes neurais (Computação); Síndrome das apneias do sono; Coração – Doenças – Diagnóstico; Python (Linguagem de programação de computador); Software – Desenvolvimento
• Link de acesso: http://hdl.handle.net/1843/50700; https://orcid.org/0000-0002-6524-7325

The application of machine learning has become increasingly common in various professional areas. Among the various possibilities that this technology allows, using it as a prediction and decision-making tool has been very promising. However, the "Black-Box" feature of some models has made the use of this technology unviable, especially in the medical field. Healthcare professionals need clarity about the factors that indicate a diagnosis. After all, a wrong diagnosis can lead to a patient’s life. Using a database with more than 560,000 records of medical appoint ments in patients, this work proposes a methodology that builds explainable machine learning models for the diagnosis prediction of auricular fibrillation, coronary sickness, and sleep apnea, incorporating patient’s structured and unstructured historical data. Weak supervision is used to label the unstructured data, XGBoost is used for prediction, and the SHAP method is used to explain the prediction. Finally, the methodology is implemented in Web Software written in the Python programming language. The results are promising, in addition to the accurate prediction capability, the prediction explanation highlights the patient’s historical characteristics with a higher impact on the decision-making process of the suggested diagnostic.