Um modelo híbrido de mineração de dados para suspeita diagnóstica relacionada a síndromes demenciais

• Autor(a) principal: Moreira, Leonard Barreto
• Data de Publicação: 2016
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UERJ
• Texto Completo: http://www.bdtd.uerj.br/handle/1/13702
• topic: Data mining; Text mining; Medical diagnosis; Machine Learning; D3M; Mineração de dados; Mineração de textos; Diagnóstico Médico; Aprendizagem de máquina; D3M; Demência - Diagnóstico; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO

The creation of intelligent computational systems has been one of the most ambitious and controversial technological challenges of our era. We can highlight the support clinical decision systems, which aim at helping experts in medical research process, thus enabling more accurate diagnoses. The medical diagnosis process is not an easy task, given the diversity of diseases, symptoms, tests, as well as the complexity of human physiology. Specifically in relation to the diagnosis of dementia, such difficulty is even greater given the lack of a specific test to determine the type of dementia, thus requiring extensive clinical research through a series of appointment, assessments and exams. In the present work is developed a hybrid model of mining involving the integrated text mining structured data mining. This model aims to assist the experts in the process of diagnosis of patients with clinical suspicion of dementia. The data used in the research come from information contained in the medical records of 3281 patients attended in the Center for Alzheimer's and Parkinson (CDAP) in Campos dos Goytacazes/RJ, Brazil, collected from the developed application. After data preprocessing, 605 records with 18 different attributes were selected for the mining stage. This step was based on the domain-driven data- mining (D3M) methodology, and text mining techniques were applied in order to structure the free text format information contained in the Patient's Past History (HPP) field, for subsequent integration into the previously mentioned set of the attributes. Later we used a set of classification algorithms based on naïve Bayes, Bayesian networks and decision trees techniques in order to obtain predictive models. Ensemble methods (Bagging, Boosting and Random Forests) were employed to improve the accuracy of the models. Further, a methodology (holdout with stratified sampling) was applied to generate equally distributed classes. All implementations of those techniques are available in the Weka (Waikato for Knowledge Environent analysis) tool. Comparisons were made between the metrics related to the accuracy of the proposed hybrid model with the model involving only structured data mining. These results showed the effectiveness of the hybrid model proposed to represent the diagnostic criteria for the pathologies of interest in this work.