A Transformer-Based Network for Estimating Blood Pressure Using Facial Videos

• Published in: IEEE sensors journal Vol. 25; no. 1; pp. 1969 - 1977
• Main Authors: Clinton Tosima Manullang, Martin, Lin, Yuan-Hsiang, Chou, Nai-Kuan
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.01.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Blood pressure; Blood pressure (BP); Calibration; Computer architecture; Contact pressure; Datasets; Deep learning; Estimation; Feature extraction; noncontact physiological measurement; Remote monitoring; remote photoplethysmography (rPPG); RGB images; Sensors; Telemedicine; transformer; Transformers; Video; Videos
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2024.3496115

Blood pressure (BP) monitoring is essential for diagnosing and managing various health conditions. While traditional contact-based methods have been effective, they can be uncomfortable for continuous or prolonged monitoring. The innovative discovery of remote photoplethysmography (rPPG) brings a new era for noncontact BP measurement. In this article, a transformer-based deep learning network named BP network (BPNet) was proposed to estimate noncontact BP from RGB videos. The BPNet comprises three primary components: the signal branch, feature branch, and predictor. The architecture is designed to integrate information from rPPG signal and their derivatives, rPPG features, and user inputs. A standout feature of our model is its capability to work without the need for calibration, making it more user-friendly. We assessed our model, BPNet, using two diverse datasets: our BESTLab dataset and the externally sourced Vital Video (VV) dataset, which is noted for its varied subject demographics and extensive BP distribution. The results show that BPNet outperforms recent benchmarks, marking a significant advancement in noncontact BP measurement technology. It also showed greater efficiency in terms of inference time and model complexity. In the future, the approach might focus on developing a fully automated deep learning system that removes the need for manual preprocessing and rPPG extraction. Furthermore, adding subject's demographic features and medical history could improve accuracy.