Flexible Time Series Matching for Clinical and Behavioral Data

• Autor(a) principal: Matias, Pedro António Correia
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/157934
• topic: Time Series; Pattern Segmentation; Anomaly Detection; Deep Learning; Domínio/Área Científica::Engenharia e Tecnologia::Outras Engenharias e Tecnologias

Time Series data became broadly applied by the research community in the last decades after a massive explosion of its availability. Nonetheless, this rise required an improvement in the existing analysis techniques which, in the medical domain, would help specialists to evaluate their patients condition. One of the key tasks in time series analysis is pattern recognition (segmentation and classification). Traditional methods typically perform subsequence matching, making use of a pattern template and a similarity metric to search for similar sequences throughout time series. However, real-world data is noisy and variable (morphological distortions), making a template-based exact matching an elementary approach. Intending to increase flexibility and generalize the pattern searching tasks across domains, this dissertation proposes two Deep Learning-based frameworks to solve pattern segmentation and anomaly detection problems. Regarding pattern segmentation, a Convolution/Deconvolution Neural Network is proposed, learning to distinguish, point-by-point, desired sub-patterns from background content within a time series. The proposed framework was validated in two use-cases: electrocardiogram (ECG) and inertial sensor-based human activity (IMU) signals. It outperformed two conventional matching techniques, being capable of notably detecting the targeted cycles even in noise-corrupted or extremely distorted signals, without using any reference template nor hand-coded similarity scores. Concerning anomaly detection, the proposed unsupervised framework uses the reconstruction ability of Variational Autoencoders and a local similarity score to identify non-labeled abnormalities. The proposal was validated in two public ECG datasets (MITBIH Arrhythmia and ECG5000), performing cardiac arrhythmia identification. Results indicated competitiveness relative to recent techniques, achieving detection AUC scores of 98.84% (ECG5000) and 93.32% (MIT-BIH Arrhythmia).