Machine learning algorithms development for sleep cycles detection and general physical activity based on biosignals

• Autor(a) principal: Braga, Rodrigo Duarte
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/152018
• topic: Deep Learning; Machine Learning; Wearable; Photoplethysmography; Sleep Stages; Heart Rate Variation; Domínio/Área Científica::Engenharia e Tecnologia::Outras Engenharias e Tecnologias

In this work, machine learning algorithms for automatic sleep cycles detection were developed. The features were selected based on the AASM manual, which is considered the gold standard for human technicians. These include features such as saturation of peripheral oxygen or others related to heart rate variation. As normally, the sleep phases naturally differ in frequency, to balance the classes within the dataset, we either oversampled the least common sleep stages or undersampled the most common, allowing for a less skewed performance favouring the most represented stages, while simultaneously improving worst-stage classification. For training the models we used MESA, a database containing 2056 full overnight unattended polysomnographies from a group of 2237 participants. With the goal of developing an algorithm that would only require a PPG device to be able to accurately predict sleep stages and quality, the main channels used from this dataset were SpO2 and PPG. Employing several popular Python libraries used for the development of machine learning and deep learning algorithms, we exhaustively explored the optimisation of the manifold parameters and hyperparameters conditioning both the training and architecture of these models in order for them to better fit our purposes. As a result of these strategies, we were able to develop a neural network model (Multilayer perceptron) with 80.50% accuracy, 0.7586 Cohen’s kappa, and 77.38% F1- score, for five sleep stages. The performance of our algorithm does not seem to be correlated with sleep quality or the number of transitional epochs in each recording, suggesting uniform performance regardless of the presence of sleep disorders. To test its performance in a different real-world scenario we compared the classifications attributed by a popular sleep stage classification android app, which collected information using a smartwatch, and our algorithm, using signals obtained from a device developed by PLUX. These algorithms displayed a strong level of agreement (90.96% agreement, 0.8663 Cohen’s kappa).