Differentially and Integrally Attentive Convolutional-based Photoplethysmography Signal Quality Classification

Rafael Lima, Italo Sandoval, Pedro Garcia Freitas, Maíssa Maniezzo, Arthur Lopes da Silva Valencio

Keywords: Differential Attention, Differential Inteh Attention, Signal Quality, Photoplethysmography, Wearables

TL;DR: Improving signal quality classification in photoplethysmography-based health applications using differential and integral attention

Abstract:

Photoplethysmography (PPG) is a non-intrusive and cost-effective optical technology that detects changes in blood volume within tissues, providing insights into the body’s physiological dynamics over time. By analyzing PPG data as a time series, valuable information about cardiovascular health and other physiological parameters such as Heart Rate Variability (HRV), Peripheral Oxygen Saturation (SpO2), and sleep status can be estimated. With the ever increasing user adoption of wearable devices like smartwatches, Health Monitoring Applications (HMA) are gaining popularity due to their ability to track various health metrics, including sleep patterns, heart rate, and activity tracking, by making use of PPG sensors to monitor different aspects of an individual’s health and wellness. However, reliable
health indicators require high-quality PPG signals, which are often contaminated with noise and artifacts caused by movement when using wearables. Hence, Signal Quality Assessment (SQA) is crucial in determining the trustworthiness of PPG data for HMA applications. We present a new PPG SQA approach, leveraging recent advancements in differential and integral attention-based strategies coupled with a two-stage procedure for promptly discarding highly anomalous segments, as a means of enhancing the performance of Convolutional Neural Network (CNN)-based SQA classifiers, balancing storage size and classifier accuracy in resulting models of increased robustness across PPG signals from different devices. Our methods are capable of achieving F1-scores between 0.9194 and 0.9865 across four expert-annotated datasets from different wearable devices.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 6230