Respiratory rate estimation applying non-negative matrix partial co-factorization from breath sounds

Respiratory rate (RR) stands as a crucial vital sign within both healthcare and home health contexts, garnering significant attention. Its ability to signal critical physiological changes renders it indispensable for continuous and non-invasive health monitoring, presenting a promising avenue for early detection of health issues. However, despite its importance, the accuracy of RR measurement and recording remains a persistent challenge in vital signs monitoring. This paper introduces two primary contributions. Firstly, it presents a newly developed labeled biomedical database, called BreathMY, comprising 1200 recordings of respiratory sounds captured using a conventional smartphone. Given the scarcity of labeled biomedical databases specifically tailored for RR analysis, this resource could be considered of interest for the development of novel approaches to accurately estimate this vital sign. Secondly, an approach based on Non-Negative Matrix Partial Co-Factorization (NMPCF) is proposed to extract temporal patterns recurring across respiratory cycles for RR estimation. Evaluation of RR estimation performance is conducted using the aforementioned database in which ambient noise typically encountered in home environments has also been mixed, across several signal-to-noise ratio (SNR). Results indicate the superior performance of the proposed NMPCF approach compared to standard Non-Negative Matrix Factorization (NMF) techniques, surpassing even a widely used state-of-the-art method across varying noise levels and RR values.