Machine learning algorithm development of SPO2 sensor for improved robustness in wearables

• Autor(a) principal: Veiga, Pedro de Bastos
• 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/10362/157124
• topic: Photoplethysmography; Pulse Oximetry; Machine Learning; Signal Quality; Heart Rate; Respiratory Rate; Domínio/Área Científica::Engenharia e Tecnologia::Outras Engenharias e Tecnologias

Wearable devices application in the digital measurement of health has gained attention by researchers. These devices allow for data acquisition during real-life activities, resulting in higher data availability. They often include photoplethysmography (PPG) sensors, the sensor behind pulse oximetry. Pulse oximetry is a non-invasive method for continuous oxygen saturation (SpO2) measurements, a standard monitor for anesthesia procedures, an essential tool for managing patients undergoing pulmonary rehabilitation and an effective method for assessing sleep-disordered breathing. However, the current market focuses on heart rate measurements and lacks the robustness of clinical applications for SpO2 assessment. In addition, the most common obstacle in PPG measurements is the signal quality, especially in the form of motion artifacts. Thus, this work aims at increasing the clinical robustness in this devices by evaluating its quality and then extracting relevant metrics. Firstly, a data acquisition protocol was developed, focused on acquiring data during daily activities. This resulted in a dataset with different signal qualities, which was manually annotated to be used as the base for the Machine Learning models. A second protocol was also developed especially designed for the extraction of the SpO2 measurement. Several Machine Learning models were developed to evaluate the signal in three distinct qualities (corrupted, suboptimal, optimal) in real time. A Random Forest classifier achieved accuracies of 79% and 80% for the binary models capable of differentiating between usable and unusable signals, and accuracies of 74% and 80% when distinguishing between optimal and suboptimal signals, for the two utilized channels. The multi-class models achieved accuracies of 66% and 65% for the two utilized channels. Three clinically relevant metrics were also extracted from the PPG signal: heart rate, respiratory rate and SpO2. The heart rate and respiratory rate algorithms resulted in performances similar to the ones found in the literature and in other devices currently on the market. However, while promising, more data is needed to reach statistical significance for the SpO2 measurement.