Risk situation detector for elderly people based on time-series analysis

• Ano de defesa: 2024
• Autor(a) principal: Much, Maicon Diogo
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Pontifícia Universidade Católica do Rio Grande do Sul; Escola Politécnica; Brasil; PUCRS; Programa de Pós-Graduação em Ciência da Computação
• Palavras-chave em Português: Elderly Risk Situation Detection; Time Series Analysis; Detecção de Situações de Risco em Idosos; Análise de Séries Temporais; CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO
• Link de acesso: https://tede2.pucrs.br/tede2/handle/tede/11672

Elderly people are more exposed to risk situations such as falls, sudden changes in vital signs and fainting. These situations become more common at this stage of life due to the natural decrease in the body’s ability to coordinate movements adequately. Numerous studies have proposed health monitoring systems for this population group by analyzing accelerometry data and/or based on machine learning algorithms, but the use of these systems in real situations has shown that this approach is still insufficient to accurately differentiate a risk situation from an elderly person’s daily activities. This project proposes the development of an effective and reliable health monitoring system for the elderly, through the continuous collection of time series extracted from movement sensors associated with vital signs. These signals feed a deep neural network architecture of the Long Short-Term Memory (LSTM) type, capable of interpreting them taking into account not only the moment of collection, but the entire context before and after the risk situation. This architecture was based on the hypothesis that there is a significant change in vital signs associated with a real fall. For this evaluation, an environment composed of a wearable device simulator, a mobile application simulator and a cloud system simulator was created, very close to the real scenario. This system, in its final model, presented an overall accuracy of 97%, showing that sensor fusion in a continuous data analysis architecture contributes to increasing the elderly risk detection capacity.