A carregar...
Publicação
Deteção de patologia cardíaca utilizando aprendizagem profunda
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 5.12 MB | Adobe PDF |
Orientador(es)
Resumo(s)
As doenças cardiovasculares representaram 33% dos óbitos registados em todo o mundo no
ano de 2019. Esta situação, juntamente com a ausência de sintomas evidentes, torna o
diagnóstico precoce dessas condições consideravelmente mais complexo do que o habitual.
Assim, anualmente são conduzidos vários estudos com o objetivo eficaz de detetar essas
doenças antes que se tornem problemáticas.
Esta dissertação tem como propósito avaliar a capacidade de construir um modelo de
aprendizagem profunda capaz de classificar os sons cardíacos de um paciente como normais
ou anormais. Para alcançar este objetivo, utilizou-se o conjunto de dados do concurso George
B. Moody PhysioNet Challenge para treinar modelos para a resolução do problema em
questão.
Foram propostas duas abordagens distintas que se distinguem pela forma de construção dos
modelos a serem utilizados. Na primeira abordagem, é criado um modelo de classificação para
cada ponto de auscultação do coração, sendo a classificação final determinada com base em
vários modelos. Já a segunda abordagem propõe a construção de um único modelo que
recebe um segmento de áudio de cada ponto de auscultação e fornece uma classificação final
do estado do paciente.
Em ambas as abordagens, foram aplicadas diferentes segmentações às amostras de áudio
utilizadas, foram extraídas caraterísticas de espetrogramas de mel e MFCC, e foram utilizados
diversos algoritmos de aprendizagem foram empregues, nomeadamente algoritmos de
aprendizagem profunda.
A melhor performance foi alcançada com a aplicação da primeira abordagem, obtendo um
overall accuracy de 77.36% e um F1-Score de 62.22%. Este estudo acaba por fundamentar um
ponto de progresso na resolução deste tipo de problema, demonstrando a viabilidade da
utilização de aprendizagem profunda no seu percurso.
Cardiovascular diseases accounted for 33% of the deaths recorded worldwide in 2019. This, coupled with the absence of evident symptoms, makes early diagnosis of these conditions considerably more complex than usual. Consequently, numerous studies are conducted annually with the effective aim of detecting these diseases before they become problematic. This dissertation aims to assess the ability to construct a deep learning model capable of classifying a patient's cardiac sounds as normal or abnormal. To achieve this goal, the dataset from the George B. Moody PhysioNet Challenge was utilized to train models for addressing the stated problem. Two distinct approaches were proposed, differing in the manner of constructing the models to be employed. In the first approach, a classification model is created for each cardiac auscultation point, and the final classification is determined based on various models. Conversely, the second approach suggests building a single model that takes an audio segment from each auscultation point and provides a final classification of the patient's condition. In both approaches, different segmentations of the audio samples were applied, features were extracted for both mel spectrograms and MFCCs, and various learning algorithms were employed, including deep learning algorithms. The best performance was achieved by applying the first approach, yielding an overall accuracy of 77.36% and an F1-Score of 62.22%. This study ultimately substantiates a point of progress in addressing this type of problem, demonstrating the viability of employing deep learning in its course.
Cardiovascular diseases accounted for 33% of the deaths recorded worldwide in 2019. This, coupled with the absence of evident symptoms, makes early diagnosis of these conditions considerably more complex than usual. Consequently, numerous studies are conducted annually with the effective aim of detecting these diseases before they become problematic. This dissertation aims to assess the ability to construct a deep learning model capable of classifying a patient's cardiac sounds as normal or abnormal. To achieve this goal, the dataset from the George B. Moody PhysioNet Challenge was utilized to train models for addressing the stated problem. Two distinct approaches were proposed, differing in the manner of constructing the models to be employed. In the first approach, a classification model is created for each cardiac auscultation point, and the final classification is determined based on various models. Conversely, the second approach suggests building a single model that takes an audio segment from each auscultation point and provides a final classification of the patient's condition. In both approaches, different segmentations of the audio samples were applied, features were extracted for both mel spectrograms and MFCCs, and various learning algorithms were employed, including deep learning algorithms. The best performance was achieved by applying the first approach, yielding an overall accuracy of 77.36% and an F1-Score of 62.22%. This study ultimately substantiates a point of progress in addressing this type of problem, demonstrating the viability of employing deep learning in its course.
Descrição
Palavras-chave
Aprendizagem Profunda CNN doenças cardiovasculares modelos pré-treinados