Graphs of growth : detecting infant movement anomalies with graph convolutional networks

• Ano de defesa: 2024
• Autor(a) principal: Sch?ler, Guilherme Gr?f
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Pontif?cia Universidade Cat?lica do Rio Grande do Sul; Escola Polit?cnica; Brasil; PUCRS; Programa de P?s-Gradua??o em Ci?ncia da Computa??o
• Palavras-chave em Português: General Movements Assessmen; Convolutional Neural Networks; Deep Learning; General Movements; Avalia??o de Movimentos Gerais; Redes Neurais Convolutivas; Aprendizado Profundo; Movimentos Gerais; CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO
• Link de acesso: https://tede2.pucrs.br/tede2/handle/tede/11353

Cognitive development disorder (CDD) is an umbrella term for impairments arising from the maldevelopment of the nervous system. Premature infants are the most affected population and although most CDDs have no cure, treatment is available as soon as the disorder is identified. The General Movements Assessment (GMA) is a diagnostic tool for discerning between typical and disorder-like neurodevelopment of infants below 6 months of age via the observation of specific movement repertoires -- some of which are abnormal and attribute risk to the infant. Despite its high predictive value for CDDs, GMA is scarcely used in clinical settings due to a difficult and costly training and certification program. This dissertation?s purpose is to develop a methodology for automating GMA: from video-recordings of moving infants in hospital settings to the classification of normal and abnormal movement and later risk identification. We developed a classification system based on a Graph Convolutional Neural Network to sort out infant skeleton time-series data of three different publicly available datasets into risk of CDDs and no-risk of CDDs. In total, data from 137 infants were used to train our classification algorithm. Changes to the internal architecture of the network and regularization steps were made to adapt to the noisy nature of our data. We performed hyperparameter optimization on different experimental setups, subjecting our model to different data, both intra-datasets ? training and testing on the same dataset and ? and inter-datasets