# The science of rPPG

> How remote photoplethysmography (rPPG) reads your vitals from a camera. The university research, clinical validation, FDA clearances and honest limits behind face scan health.

rPPG recovers the blood volume pulse from tiny color changes in your skin captured on ordinary video. The field runs from the 2008 origin work through recent transformer and state space models and fresh clinical validations. Vital Health Scan itself is not FDA cleared. The clearances below describe the underlying rPPG technology as cleared in other companies' devices.

## Research

- **Remote plethysmographic imaging using ambient light** (Verkruysse, Svaasand & Nelson, Optics Express 16(26), 2008) — Proven, Beckman Laser Institute, UC Irvine. The origin paper. It proved a plain consumer camera under ordinary room light can recover the blood volume pulse from skin at over a meter away, with the green channel carrying the strongest signal. [Source](https://pubmed.ncbi.nlm.nih.gov/19104573/)
- **Non contact, automated cardiac pulse measurements using video imaging and blind source separation** (Poh, McDuff & Picard, Optics Express 18(10), 2010) — Proven, MIT Media Lab. The webcam breakthrough. Heart rate from ordinary video agreed with an FDA approved finger sensor at correlation above 0.98 in controlled conditions, even with several people in frame. [Source](https://pubmed.ncbi.nlm.nih.gov/20588929/)
- **Advancements in noncontact, multiparameter physiological measurements using a webcam** (Poh, McDuff & Picard, IEEE Trans. Biomedical Engineering 58(1), 2011) — Proven, MIT Media Lab. Showed a single webcam video can recover heart rate, respiratory rate AND heart rate variability at once, all validated against FDA approved sensors. [Source](https://pubmed.ncbi.nlm.nih.gov/20952328/)
- **Robust pulse rate from chrominance based rPPG** (de Haan & Jeanne, IEEE Trans. Biomedical Engineering 60(10), 2013) — Proven, Philips Research & TU Eindhoven. Introduced the CHROM method to cancel motion artefacts. On 117 subjects it hit about 92% agreement with contact PPG, and under exercise correct pulse detection rose from 79% to 98%. [Source](https://research.tue.nl/en/publications/robust-pulse-rate-from-chrominance-based-rppg)
- **Algorithmic principles of remote PPG** (Wang, den Brinker, Stuijk & de Haan, IEEE Trans. Biomedical Engineering 64(7), 2017) — Proven, Philips Research & TU Eindhoven. Derived the Plane Orthogonal to Skin (POS) method from a physical model of skin light reflection. It is now a standard high accuracy, motion robust baseline across the field. [Source](https://pubmed.ncbi.nlm.nih.gov/28113245/)
- **DeepPhys: video based physiological measurement using convolutional attention networks** (Chen & McDuff, ECCV 2018, 2018) — Proven, Microsoft Research & MIT Media Lab. The first end to end deep learning model to recover pulse and breathing from video, outperforming classic methods and launching the modern AI generation of rPPG. [Source](https://arxiv.org/abs/1805.07888)
- **PhysFormer: facial video based physiological measurement with temporal difference transformer** (Zitong Yu et al., CVPR 2022 / Int. Journal of Computer Vision 2023, 2022) — Proven, Tsinghua University & University of Oxford. Brought the transformer architecture to face scanning, reaching heart rate error near 1.1 bpm on the hard VIPL-HR benchmark and about 0.4 bpm on PURE while training from scratch on small datasets. [Source](https://arxiv.org/abs/2302.03548)
- **Contrast-Phys+: unsupervised and weakly supervised remote physiological measurement via spatiotemporal contrast** (Zhaodong Sun & Xiaobai Li, IEEE Trans. Pattern Analysis and Machine Intelligence (TPAMI), 2024) — Proven, University of Oulu. Learns to read the pulse with no ground truth labels using a contrastive loss. Across five public datasets it beat prior label free methods and approached the best supervised models. [Source](https://arxiv.org/abs/2309.06924)
- **RhythmMamba: fast, lightweight, and accurate remote physiological measurement** (Zou, Guo, Hu & Ma, arXiv (cs.CV), 2024) — Proven, University of Science and Technology Beijing. A modern state space model that reads any length of video with tiny compute. Heart rate error of 0.23 bpm on PURE and 0.50 bpm on UBFC-rPPG, with strong cross dataset transfer, using about one million parameters. [Source](https://arxiv.org/abs/2404.06483)
- **Advancing generalizable remote physiological measurement through explicit and implicit prior knowledge** (Zhang, Lu, Liu, Chen & Wu, arXiv (cs.CV), 2024) — Proven, HKUST (Guangzhou). Combines known physiology with learned priors so a model trained on one dataset holds up on unseen ones, improving robustness across cameras, lighting and even from color to infrared video. [Source](https://arxiv.org/abs/2403.06947)
- **Feasibility of assessing ultra short term pulse rate variability from video recordings** (Sensors, Sensors / PMC, 2020) — Proven, Peer reviewed validation (Sensors). From 60 second facial videos, camera HRV correlated with reference PPG (SDNN r = 0.84), confirming that video HRV is feasible. [Source](https://pmc.ncbi.nlm.nih.gov/articles/PMC6953345/)
- **Non contact physiological monitoring of preterm infants in the neonatal intensive care unit** (Villarroel, Jorge, Tarassenko et al., npj Digital Medicine 2:128, 2019) — Proven, University of Oxford & John Radcliffe Hospital. Across 426 hours of video from 30 preterm infants, camera heart rate reached a mean error of 2.3 bpm and respiratory rate 3.5 breaths per minute for the majority of the time. [Source](https://pmc.ncbi.nlm.nih.gov/articles/PMC6908711/)
- **Non contact physiological monitoring of post operative patients in the intensive care unit** (Jorge, Villarroel, Tarassenko et al., npj Digital Medicine 5:4, 2022) — Proven, University of Oxford, Institute of Biomedical Engineering. In adult ICU patients, camera heart rate matched wired monitors with a mean error of 2.5 bpm (r about 0.98) and respiratory rate with a mean error of 2.4 breaths per minute. [Source](https://pmc.ncbi.nlm.nih.gov/articles/PMC8758749/)
- **Contactless facial video recording with deep learning for the detection of atrial fibrillation** (Sun, Yang, Wu et al., Scientific Reports 12:281, 2022) — Proven, National Yang Ming Chiao Tung University & partner hospitals. In 453 patients, facial video screened for atrial fibrillation with 90% accuracy on 30 second clips and 97% on longer recordings. [Source](https://pmc.ncbi.nlm.nih.gov/articles/PMC8741942/)
- **Remote photoplethysmography is an accurate method to remotely measure respiratory rate** (Frontiers in Physiology, Frontiers in Physiology, 2022) — Proven, Hospital based trial (963 patients). In 963 hospital patients, camera respiratory rate showed 96.0% agreement with the standard clinical method, one of the largest validations of its kind. [Source](https://pmc.ncbi.nlm.nih.gov/articles/PMC9267568/)
- **Clinical validation of rPPG enabled contactless pulse rate monitoring in cardiovascular disease patients** (Chin, Chan et al., Bioengineering (MDPI), 2026) — Proven, PanopticAI & Prince of Wales Hospital / CUHK. In 47 cardiovascular patients over 782 measurements, camera pulse rate matched ECG with a mean error of 1.06 bpm and correlation of 0.962. The software is FDA cleared (K240890). [Source](https://pmc.ncbi.nlm.nih.gov/articles/PMC12938449/)
- **Seconds matter: rapid non contact monitoring of heart and respiratory rate from face videos** (Khan et al., Sensors (MDPI), 2026) — Proven, University of Gothenburg & Detectivio. In 200 participants (mean age 62), 15 second face videos gave heart rate within 3.25 bpm (98.5% within 10 bpm) and respiratory rate within 0.69 breaths per minute. [Source](https://pmc.ncbi.nlm.nih.gov/articles/PMC12987125/)
- **Continuous non contact vital sign monitoring of neonates in intensive care using RGB-D cameras** (Estevez et al., Scientific Reports (Nature), 2025) — Proven, University of Cambridge & Rosie Hospital NHS. A real neonatal ICU study over 744 minutes of video estimated heart rate, oxygen saturation and respiratory rate continuously against bedside monitors, an honest look at accuracy in a demanding clinical setting. [Source](https://pmc.ncbi.nlm.nih.gov/articles/PMC12081704/)
- **Advancing rPPG to facilitate cardiac monitoring in naturalistic settings using webcam technology** (Woelk et al., Behavior Research Methods (Springer), 2026) — Proven, University College London. Using consumer webcams on 77 people, heart rate error was 1.67 bpm and HRV (SDNN and RMSSD) was recovered within about 11 ms, showing HRV from an everyday webcam is within reach. [Source](https://pmc.ncbi.nlm.nih.gov/articles/PMC13106271/)
- **Exploring contactless vital signs collection in video telehealth visits** (Garvin et al., JMIR Formative Research, 2024) — Proven, Veterans Affairs Boston Healthcare System. A live deployment of camera vitals inside the VA telehealth platform with 20 providers and 13 patients scored highly for usability (86/100) and acceptability (90/100) in real appointments. [Source](https://formative.jmir.org/2024/1/e60491)
- **Estimating heart rate variability using facial video photoplethysmography: a pilot validation** (Pstras et al., medRxiv preprint, 2025) — Emerging, Polish Academy of Sciences & Wroclaw Medical University. A 35 person pilot (preprint) found smartphone video HRV tracked ECG well for overall variability (SDNN error 3.5 ms, r = 0.84) while beat to beat measures were looser, so we treat video HRV as emerging. [Source](https://www.medrxiv.org/content/10.1101/2025.02.13.25322028v1)
- **Accuracy of heart rate, oxygen saturation and blood pressure using a non contact PPG mobile app** (Zuccotti et al., Digital Health (SAGE), 2025) — Emerging, University of Milano & Buzzi Children's Hospital. In 562 participants, camera oxygen saturation reached 2.10% mean error and heart rate 2.96 bpm. Blood pressure was much weaker, an honest counterpoint we carry through our own labeling. [Source](https://pmc.ncbi.nlm.nih.gov/articles/PMC12227916/)
- **Contactless blood oxygen saturation estimation from facial videos using deep learning** (Bioengineering (MDPI), Bioengineering 11(3):251, 2024) — Emerging, Imaging pulse oximetry study (Bioengineering). A deep learning model reached 1.27% mean error, beating the plus or minus 4% accuracy standard for approved pulse oximeters under study conditions. Promising, still maturing. [Source](https://www.mdpi.com/2306-5354/11/3/251)
- **A hybrid CNN spectral architecture for non contact respiratory rate estimation** (Srestha et al., PLoS One, 2026) — Proven, Yeungnam University. Across a demographically diverse group, camera respiratory rate stayed within about 0.6 to 0.95 breaths per minute, holding up across different skin tones. [Source](https://pmc.ncbi.nlm.nih.gov/articles/PMC12923011/)
- **Enhancing stress detection through rPPG analysis and deep learning** (Fontes et al., Sensors (MDPI), 2024) — Emerging, Nottingham Trent University. Deep learning models read stress from camera pulse signals with up to 95.8% classification accuracy on a public benchmark, pointing to camera based stress screening. [Source](https://pmc.ncbi.nlm.nih.gov/articles/PMC10892284/)
- **Contactless and calibration free blood pressure and pulse rate monitor for hypertension screening** (Kapoor, Holman & Cohen, JMIR Cardio, 2024) — Experimental, Lifelight (Mind over Matter Medtech) & Element Materials. An independent lab test of a camera monitor: pulse rate was excellent (1.1 bpm error) but blood pressure did not meet the strict clinical cuff standard, so we present camera blood pressure as screening grade only. [Source](https://pmc.ncbi.nlm.nih.gov/articles/PMC11333865/)
- **Robust blood pressure measurement from facial videos in diverse environments** (Park & Hong, Heliyon (Cell Press), 2024) — Experimental, Sungkyunkwan University. Camera blood pressure held to about 4 mmHg error indoors across 520 recordings, degrading outdoors and in motion. Promising research, but not yet a clinical grade cuff replacement. [Source](https://pmc.ncbi.nlm.nih.gov/articles/PMC10906170/)
- **Clinical applications of contactless photoplethysmography for monitoring in adults** (Bautista et al., Journal of Clinical and Translational Science, 2023) — Proven, University of Leeds. A systematic review and meta-analysis of 12 studies found an overall bias of just 0.13 bpm between contact and contactless heart rate, with respiratory rate within 2 breaths 92% of the time. [Source](https://pmc.ncbi.nlm.nih.gov/articles/PMC10260340/)
- **Challenges and prospects of visual contactless physiological monitoring in clinical study** (npj Digital Medicine, npj Digital Medicine 6:231, 2023) — Proven, npj Digital Medicine (Nature Portfolio). A Nature family review synthesizing the clinical evidence, algorithms and datasets behind camera vitals, the authoritative state of the science reference for the field. [Source](https://doi.org/10.1038/s41746-023-00973-x)
- **The role of face regions in remote photoplethysmography for contactless heart rate monitoring** (npj Digital Medicine, npj Digital Medicine, 2025) — Proven, npj Digital Medicine (Nature Portfolio). Reviewing 70 studies, this paper found forehead and cheek regions give the best accuracy, with best case heart rate error around 1.0 bpm. [Source](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12297079/)
- **rPPG-Toolbox: deep remote PPG toolbox** (Xin Liu et al., NeurIPS 2023 (Datasets & Benchmarks), 2023) — Proven, University of Washington & Microsoft. An open source platform that benchmarks the leading neural and unsupervised rPPG methods on standard datasets with unified evaluation, now the reproducibility standard for the field. [Source](https://arxiv.org/abs/2210.00716)
- **Demographic bias in public remote photoplethysmography datasets** (Bondarenko, Menon & Elgendi, npj Digital Medicine (Nature Portfolio), 2025) — Proven, ETH Zurich & Khalifa University. An audit across 100 rPPG studies found older chrominance methods lose accuracy on darker skin, while modern deep learning models narrow the gap substantially though they do not fully erase it. We cite this because honesty on fairness matters. [Source](https://www.nature.com/articles/s41746-025-01973-9)
- **The reliability of remote photoplethysmography under low illumination and elevated heart rates** (Acharya, Saakyan, Hammer & Drimalla, npj Digital Medicine 8:744, 2025) — Proven, Bielefeld University. A rigorous benchmark of 8 algorithms found the best method holds near 1.1 bpm at rest, while accuracy is harder at very high heart rates, exactly the kind of honest boundary we build our labeling around. [Source](https://pmc.ncbi.nlm.nih.gov/articles/PMC12678791/)

## FDA clearances of the underlying technology

- **FaceHeart — FH Vitals SDK-RR**: FDA 510(k) K243966. Contactless video respiratory rate measurement (Class II software as a medical device). [Source](https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/pmn.cfm?ID=K243966)
- **Presage Technologies — SmartSpectra Vital Signs Monitor 1.0 SDK**: FDA 510(k) K254169. Contactless pulse rate and breathing rate via mobile camera on iOS and Android. [Source](https://www.prnewswire.com/news-releases/presage-technologies-earns-fda-clearance-for-contactless-vital-signs-makes-them-available-for-free-302818692.html)
- **PanopticAI — Vitals monitoring software**: FDA 510(k) K240890. Contactless camera pulse rate, clinically validated in cardiovascular patients. [Source](https://pmc.ncbi.nlm.nih.gov/articles/PMC12938449/)

See also: [Home](https://scanner.vitalhealthglobal.com/index.md) · [All biomarkers](https://scanner.vitalhealthglobal.com/biomarkers.md)