A1 Refereed original research article in a scientific journal
Respiration Rate Estimation via Smartwatch-based Photoplethysmography and Accelerometer Data: A Transfer Learning Approach
Authors: Kazemi, Kianoosh; Azimi, Iman; Liljeberg, Pasi; Rahmani, Amir M.
Publisher: Association for Computing Machinery
Publishing place: NEW YORK
Publication year: 2025
Journal: Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies
Journal name in source: PROCEEDINGS OF THE ACM ON INTERACTIVE MOBILE WEARABLE AND UBIQUITOUS TECHNOLOGIES-IMWUT
Journal acronym: PROC ACM INTERACT MO
Article number: 7
Volume: 9
Issue: 1
Number of pages: 24
eISSN: 2474-9567
DOI: https://doi.org/10.1145/3712280
Web address : https://doi.org/10.1145/3712280
Self-archived copy’s web address: https://research.utu.fi/converis/portal/detail/Publication/491741756
Respiration Rate (RR) is a biomarker for several illnesses that can be extracted from biosignals, such as photoplethysmogram (PPG) and accelerometers. Smartwatch-based PPG signals are more prone to noise interference, particularly within their lower frequency spectrum where respiratory data is embedded. Therefore, existing methods are insufficient for extracting RR from PPG data collected from wrists reliably. Additionally, accelerometer sensors embedded in smartwatches capture respiration-induced motion and can be integrated with PPG signals to improve RR extraction. This paper proposes a deep learning-based model to extract RR from raw PPG and accelerometer signals captured via a smartwatch. The proposed network combines dilated residual inception module and Multi-Scale convolutions. We propose a pre-trained foundation model for smartwatch-based RR extraction and apply a transfer learning technique to enhance the generalizability of our method across different datasets. We test the proposed method using two public datasets (i.e., WESAD and PPG-DaLiA). The proposed method shows the Mean Absolute Error (MAE) of 2.29 and 3.09 and Root Mean Squared Errors (RMSE) of 3.11 and 3.79 across PPG-DaLiA and WESAD datasets, respectively. In contrast, the best results obtained by the existing methods are an MAE of 2.68, an RMSE of 3.5 for PPG-DaLiA, an MAE of 3.46, and an RMSE of 4.02 for WESAD datasets.
Downloadable publication This is an electronic reprint of the original article. |  |
Funding information in the publication:
This work was partially supported by the Finnish Foundation for Technology Promotion and the Nokia Foundation.
