Diagnóstico da COVID-19 utilizando Redes Neurais Artificiais baseadas na Teoria da Ressonância Adaptativa
Carregando...
Arquivos
Data
2023-01-27
Autores
Silva, Reginaldo José da
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
A COVID-19 espalhou-se por todo o mundo levando a óbito cerca de 6,6 milhões de pessoas, diversas estratégias foram utilizadas para evitar a propagação dessas infecções, estratégias essas que influenciaram a vida de diversas pessoas. A forma mais eficaz para controle da doença é a detecção e o isolamento que limita a propagação do vírus, consequentemente, salvando vidas. Dentre os diversos testes existentes para detecção e triagem da COVID-19, o teste RT-PCR é considerado o padrão-ouro. No entanto, o teste possui baixa sensibilidade de diagnóstico, resultando numa taxa elevada de falsos negativos. Diversos estudos clínicos mostraram alterações hematológicas e bioquímicas em pacientes infectados, e diante da necessidade por diagnósticos eficientes e mais precisos em fases inicias, surgiram estudos que utilizam estas alterações com algoritmos de aprendizagem de máquinas para ajudar a diagnosticar a doença. Desta forma, o presente trabalho tem por objetivo avaliar o desempenho de diferentes redes neurais baseadas na Teoria da Ressonância Adaptativa (ART), bem como desenvolver uma metodologia rápida e eficaz para a classificação da COVID-19 utilizando exames laboratoriais. Diversas redes ART foram testadas, dentre elas autoexpansíveis e com treinamento continuado, vale ressaltar que as redes ART autoexpansíveis realizam o processo de classificação de forma mais rápida que as redes ART tradicionais, enquanto as redes com treinamento continuado permitem a aprendizagem de modo on-line, tornando o treinamento e a análise mais eficazes. Um total de doze conjuntos de dados de cinco países diferentes contendo diversos biomarcadores sanguíneos foram utilizados, a metodologia foi dividida em duas etapas, a primeira responsável por avaliar a melhor rede ART para estes dados e a segunda por avaliar o comportamento desta rede na classificação de pacientes com COVID-19 em dados de diferentes países. A rede ART Fuzzy Autoexpansível com Treinamento Continuado obteve o melhor desempenho, superando na maior parte das vezes, modelos tradicionais e de aprendizado profundo frequentemente utilizados pela literatura. Além disso, os atributos selecionados mostraram excelente capacidade preditiva ao classificar com eficiência pacientes de diferentes países. Logo, a metodologia proposta pode servir como um método complementar ao RT-PCR, diminuindo a taxa de falsos positivos ou negativos, ou até mesmo como uma alternativa de baixo custo para a detecção.
COVID-19 has spread all over the world, killing about 6.6 million people. Several strategies have been used to prevent the spread of these infections, strategies that have influenced the lives of many people. The most effective way to control the disease is the detection and isolation that limits the spread of the virus, consequently saving lives. Among the various existing tests for detection and screening of COVID-19, the RT-PCR test is considered the gold standard. However, the test has low diagnostic sensitivity, resulting in a high false negative rate. Several clinical studies have shown hematological and biochemical alterations in infected patients, and facing the need for efficient and more accurate diagnosis in early stages, studies have emerged that use these alterations with machine learning algorithms to help diagnose the disease. Thus, the present work aims to evaluate the performance of different neural networks based on Adaptive Resonance Theory (ART), as well as to develop a fast and effective methodology for the classification of COVID-19 using laboratory tests. Several ART networks were tested, including self-expanding and continuously trained networks; it is worth noting that self-expanding ART networks perform the classification process faster than traditional ART networks, while continuously trained networks allow learning in an online manner, making training and analysis more effective. A total of twelve datasets from five different countries containing various blood biomarkers were used, the methodology was divided into two stages, the first responsible for evaluating the best ART network for these data and the second for evaluating the behavior of this network in classifying patients with COVID-19 in data from different countries. The Fuzzy Self-expandable ART network with Continuous Training obtained the best performance, outperforming for the most part, traditional and deep learning models often used by the literature. Furthermore, the selected attributes showed excellent predictive ability by efficiently classifying patients from different countries. Therefore, the proposed methodology can serve as a complementary method to RT-PCR, decreasing the false positive or negative rate, or even as a low cost alternative for detection.
COVID-19 has spread all over the world, killing about 6.6 million people. Several strategies have been used to prevent the spread of these infections, strategies that have influenced the lives of many people. The most effective way to control the disease is the detection and isolation that limits the spread of the virus, consequently saving lives. Among the various existing tests for detection and screening of COVID-19, the RT-PCR test is considered the gold standard. However, the test has low diagnostic sensitivity, resulting in a high false negative rate. Several clinical studies have shown hematological and biochemical alterations in infected patients, and facing the need for efficient and more accurate diagnosis in early stages, studies have emerged that use these alterations with machine learning algorithms to help diagnose the disease. Thus, the present work aims to evaluate the performance of different neural networks based on Adaptive Resonance Theory (ART), as well as to develop a fast and effective methodology for the classification of COVID-19 using laboratory tests. Several ART networks were tested, including self-expanding and continuously trained networks; it is worth noting that self-expanding ART networks perform the classification process faster than traditional ART networks, while continuously trained networks allow learning in an online manner, making training and analysis more effective. A total of twelve datasets from five different countries containing various blood biomarkers were used, the methodology was divided into two stages, the first responsible for evaluating the best ART network for these data and the second for evaluating the behavior of this network in classifying patients with COVID-19 in data from different countries. The Fuzzy Self-expandable ART network with Continuous Training obtained the best performance, outperforming for the most part, traditional and deep learning models often used by the literature. Furthermore, the selected attributes showed excellent predictive ability by efficiently classifying patients from different countries. Therefore, the proposed methodology can serve as a complementary method to RT-PCR, decreasing the false positive or negative rate, or even as a low cost alternative for detection.
Descrição
Palavras-chave
ART Fuzzy, Biomarcadores sanguíneos, Classificação, Reconhecimento de padrões, SARS-CoV-2, Treinamento Continuado, Fuzzy ART, Blood biomarkers, Classification, Pattern recognition, Continuous Training