Detecção de alterações respiratórias na esclerose sistêmica através da técnica de oscilações forçadas e algoritmos de aprendizado de máquinas

• Autor(a) principal: Andrade, Domingos Sávio Mattos de
• Data de Publicação: 2018
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UERJ
• Texto Completo: http://www.bdtd.uerj.br/handle/1/11751
• topic: Electronic Engineering; Computer learning; Algorithms; Respiratory system diseases; Systemic Sclerosis; Forced Oscillations Technique; TOF; Machine learning; Multiple Instances; AUC; Engenharia eletrônica; Aprendizado do computador; Algorítmos; Doenças do sistema respiratório; Esclerose Sistêmica; Técnica de Oscilações Forçadas; TOF; Aprendizado de máquina; Múltiplas Instâncias; AUC; CNPQ::ENGENHARIAS

The Forced Oscillation Technique (TOF) has shown high potential in the study of mechanical alterations of respiratory diseases. Due to its simplicity of execution, requiring little patient´s collaboration, it provides new parameters, which complement the pulmonary evaluations performed by traditional methods, such as spirometry. To prove the efficiency of this new technique, this work proposes the use of machine learning algorithms help the investigation and the early diagnosis of respiratory changes in patients with Systemic Sclerosis. The data provided by TOF were used on the following algorithms: K-Nearest Neighbor (KNN), Radial Support Vector Machine (RSVM), Adaboost (ADAB), Random Forest (RF) and Multiple Instance Learning (AMI) as classifier and attribute selector. From the experiments performed, dynamic compliance (Cdin) was the attribute that presented the best individual performance (AUC = 0,77 for the control group x normal spirometric examination (group 1) and 0,94 for the control group x restrictive (group 2)), indicating moderate and high accuracy, respectively. In the experiment with all TOF attributes, the best performance was achieved by KNN (AUC = 0,90) for group 1, indicating high accuracy and AUC = 0,96, achieved by ADAB for group 2. The AMI algorithm obtained AUC = 0,72 for group 1 and AUC = 0,94 for group 2. With five attributes selected by the multi-instance algorithm, the best result was achieved by the KNN algorithm (AUC = 0,88) for group 1 and ADAB (AUC = 0,98) for group 2. As a classifier, the AMI obtained AUC of 0,74 and 0,95, respectively. In the experiment performed with five attributes selected by the expert, the best performance was achieved by KNN (AUC = 0,85) for group 1 and ADAB (AUC = 0,98) for group 2. With direct selection, the best performance was also achieved by KNN for group 1 (AUC = 0,89) and by ADAB for group 2 (AUC = 0,97). The experiments showed that the use of machine learning algorithms increased the accuracy of the diagnosis of Systemic Sclerosis, which may help in the early diagnosis of the disease.