Estimation of Blood Pressure Using Remote Photoplethysmography and Deep Learning
Hantula, Olli (2025)
2025
Bioteknologian ja biolääketieteen tekniikan maisteriohjelma - Master's Programme in Biotechnology and Biomedical Engineering
Lääketieteen ja terveysteknologian tiedekunta - Faculty of Medicine and Health Technology
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2025-08-18
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202508188300
https://urn.fi/URN:NBN:fi:tuni-202508188300
Tiivistelmä
High blood pressure, or hypertension, is the leading preventable cause of premature death globally, yet nearly half of those affected remain undiagnosed. Traditional blood pressure measurement methods, such as cuff-based devices, can be inconvenient, costly, or inaccurate due to the stress of visiting healthcare facilities – a phenomenon known as “white coat hyper-tension”. This creates a pressing need for more accessible, continuous, and user-friendly alternatives.
This work builds upon a previously published method by Hamoud et al. [1], which uses convolutional neural networks and long short-term memory (LSTM) networks to estimate systolic and diastolic blood pressure from facial cheek patches. The original approach is reimplemented and trained with a larger dataset, and its performance is critically evaluated. The method is further extended by integrating remote photoplethysmography – a contactless, camera-based technique that detects blood volume pulse signals via subtle skin color changes caused by blood flow. Additionally, the impact of sequence length (1 vs. 8 seconds) on model performance is investigated.
The experimental setup involved training eight deep learning models using the Vision for Vitals (V4V) dataset [2], which includes synchronized facial videos and continuous blood pressure recordings. The models were evaluated using multiple metrics, including mean absolute error, mean accuracy, and Pearson’s correlation coefficient. Despite rigorous training and pre-processing, the models exhibited significant overfitting and failed to outperform simple base-line estimators, raising concerns about their generalizability and clinical applicability.
The significance of this research lies not only in its technical contributions but also in its critical examination of current practices in machine learning-based health monitoring. By high-lighting the limitations of existing models and emphasizing the need for robust evaluation standards, this work contributes to the development of more reliable and transparent AI systems in healthcare.
Contactless blood pressure monitoring via smartphones or webcams could revolutionize public health by enabling widespread, low-cost, and continuous cardiovascular monitoring. This would be particularly impactful in low-resource settings and for populations with limited access to healthcare.
[1] B. Hamoud, A. Kashevnik, W. Othman, and N. Shilov, ‘Neural Network Model Combination for Video-Based Blood Pressure Estimation: New Approach and Evaluation’, Sensors, vol. 23, no. 4, Art. no. 4, Jan. 2023, doi: 10.3390/s23041753.
[2] A. Revanur, Z. Li, U. A. Ciftci, L. Yin, and L. A. Jeni, ‘The First Vision For Vitals (V4V) Challenge for Non-Contact Video Based Physiological Estimation’, Sep. 22, 2021, arXiv: arXiv:2109.10471. doi: 10.48550/arXiv.2109.10471.
This work builds upon a previously published method by Hamoud et al. [1], which uses convolutional neural networks and long short-term memory (LSTM) networks to estimate systolic and diastolic blood pressure from facial cheek patches. The original approach is reimplemented and trained with a larger dataset, and its performance is critically evaluated. The method is further extended by integrating remote photoplethysmography – a contactless, camera-based technique that detects blood volume pulse signals via subtle skin color changes caused by blood flow. Additionally, the impact of sequence length (1 vs. 8 seconds) on model performance is investigated.
The experimental setup involved training eight deep learning models using the Vision for Vitals (V4V) dataset [2], which includes synchronized facial videos and continuous blood pressure recordings. The models were evaluated using multiple metrics, including mean absolute error, mean accuracy, and Pearson’s correlation coefficient. Despite rigorous training and pre-processing, the models exhibited significant overfitting and failed to outperform simple base-line estimators, raising concerns about their generalizability and clinical applicability.
The significance of this research lies not only in its technical contributions but also in its critical examination of current practices in machine learning-based health monitoring. By high-lighting the limitations of existing models and emphasizing the need for robust evaluation standards, this work contributes to the development of more reliable and transparent AI systems in healthcare.
Contactless blood pressure monitoring via smartphones or webcams could revolutionize public health by enabling widespread, low-cost, and continuous cardiovascular monitoring. This would be particularly impactful in low-resource settings and for populations with limited access to healthcare.
[1] B. Hamoud, A. Kashevnik, W. Othman, and N. Shilov, ‘Neural Network Model Combination for Video-Based Blood Pressure Estimation: New Approach and Evaluation’, Sensors, vol. 23, no. 4, Art. no. 4, Jan. 2023, doi: 10.3390/s23041753.
[2] A. Revanur, Z. Li, U. A. Ciftci, L. Yin, and L. A. Jeni, ‘The First Vision For Vitals (V4V) Challenge for Non-Contact Video Based Physiological Estimation’, Sep. 22, 2021, arXiv: arXiv:2109.10471. doi: 10.48550/arXiv.2109.10471.