Blood oxygen monitoring used to be a stationary affair. Clip a sensor to your finger, sit still, wait for the reading. That works fine in a clinic. It falls apart the moment someone starts running on a treadmill, cycling through intervals, or recovering between sets. And those are precisely the moments when oxygen saturation data becomes most interesting.
Exercise-induced hypoxemia — where SpO2 drops meaningfully during intense physical effort — affects a meaningful subset of endurance athletes. It was first described by Dempsey et al. in the 1980s and has been documented with SpO2 values falling between 87% and 94% during maximal exertion (Powers et al., 1984; Sheel et al., 2001). Catching those drops in real time matters for training decisions. The problem is that most monitoring methods require physical contact, and contact sensors move, slip, or compress during vigorous activity.
"Exercise-induced hypoxemia is well-described in endurance-trained athletes during both maximal and submaximal exercise intensities, yet its consequences have not been fully characterized." — Durand & Raberin, Sports Medicine (2021)
Why Exercise Makes Contactless SpO2 Harder
Camera-based SpO2 estimation is one of the harder applications of rPPG. The signal that encodes oxygen information (the differential absorption of oxygenated versus deoxygenated hemoglobin across RGB channels) is subtler than the heart rate signal that most rPPG systems extract with relative ease. Add physical movement and the difficulties compound quickly.
The core challenges break down along predictable lines:
- Motion artifacts dominate the signal. Head movement, facial muscle contraction, and postural shifts during exercise create noise that can overwhelm the pulsatile component of the rPPG signal. Van Gastel et al. (2016) at TU Eindhoven developed motion-robust algorithms for seated subjects with minor head movement, but vigorous exercise represents a different order of magnitude.
- Skin perfusion changes during exercise. Blood redistributes toward working muscles and the skin surface during physical exertion, altering the baseline optical properties of facial tissue. This changes the calibration relationship between camera-derived color ratios and actual SpO2.
- Breathing artifacts intensify. Heavy breathing during exercise creates periodic facial movements and changes in venous return that add low-frequency noise to the rPPG signal, complicating the extraction of the arterial pulsatile component.
- Sweating alters skin reflectance. A layer of sweat on the face changes how light interacts with the skin surface, introducing specular reflections that don't carry physiological information.
None of these are insurmountable. But they explain why most published rPPG SpO2 studies to date have focused on seated, resting subjects — that is where the signal-to-noise ratio cooperates.
SpO2 Monitoring Methods Compared for Exercise Scenarios
| Method | Contact Required | Motion Tolerance | Exercise Suitability | Typical Accuracy | Cost |
|---|---|---|---|---|---|
| Finger Pulse Oximeter | Yes (finger clip) | Low — dislodges easily | Poor during movement | ±2% SpO2 | $15-50 |
| Forehead Reflectance Sensor | Yes (adhesive patch) | Moderate | Fair for steady-state | ±2-3% SpO2 | $200+ (clinical) |
| Chest Strap with SpO2 | Yes (strap + sensor) | Good | Good for running/cycling | ±3% SpO2 | $100-300 |
| Smartwatch PPG | Yes (wrist contact) | Moderate — wrist motion | Fair, degrades with arm swing | ±3-5% SpO2 | $200-500 |
| rPPG Camera-Based (rest) | No | Low — requires stillness | Poor (current generation) | ±2-5% SpO2 | Camera only |
| rPPG Camera-Based (emerging) | No | Improving | Under development | TBD | Camera only |
Sources: Van Gastel et al. (2016), Verkruysse et al. (2008), FDA device databases, manufacturer specifications.
The gap in this table is obvious. No current approach handles the combination of contactless measurement and vigorous movement particularly well. That gap is exactly where the research is heading.
Research Progress on Motion-Robust Camera SpO2
Several research groups are tackling the motion problem from different angles.
Van Gastel, Stuijk, and de Haan (2016) at TU Eindhoven published foundational work on motion-robust SpO2 estimation using their Adaptive PBV (Pulse Blood Volume) method. Their approach uses the known spectral relationship between blood volume changes and SpO2 levels across RGB channels, using multiple regions of interest to separate motion artifacts from physiological signals. The work demonstrated feasibility with minor movement, establishing a baseline that subsequent researchers have built upon.
Peng et al. (2024) proposed a contrastive learning spatiotemporal attention network — a semi-supervised deep learning approach for video-based SpO2 estimation. The model learns to distinguish between motion-related and physiological signal components without requiring fully labeled datasets, which is significant because ground-truth SpO2 data during exercise is expensive to collect.
Hu et al. (2024) introduced Residual and Coordinate Attention (RCA) modules into the SpO2 estimation pipeline, using spatial attention to weight facial regions by signal quality. During movement, some facial regions maintain a cleaner signal than others — the forehead tends to be more stable than the cheeks, for instance — and attention mechanisms can automatically identify and prioritize those regions.
Cheng et al. (2024) at Hong Kong University of Science and Technology published work on contactless SpO2 estimation from facial videos, investigating the effects of different facial regions and ambient conditions on accuracy. Their findings reinforced that region selection is critical and that multi-region fusion outperforms single-region approaches.
Van Gastel et al. (2022) conducted hypoxia lab experiments with their camera-based SpO2 system, achieving errors smaller than 4 percentage points for realistic screening scenarios where subjects were seated. While not an exercise study, the controlled desaturation protocol (SpO2 range 70-100%) validated that camera-based systems can track clinically significant drops in oxygen saturation — the same kind of drops that EIH athletes experience.
Exercise-Induced Hypoxemia: The Clinical Case for Monitoring
Contactless SpO2 during exercise is not just a convenience play. Exercise-induced hypoxemia is a real phenomenon that most athletes and coaches still don't monitor for.
Prevalence and Impact
Dempsey et al. (2020) note that EIH results from an "underpowered" respiratory system in endurance athletes — the lungs simply cannot keep up with the metabolic demand during peak effort. It is more common than most people realize: estimates suggest that over 50% of highly trained endurance athletes experience some degree of arterial oxygen desaturation during maximal exercise (Prefaut et al., 2000).
This matters more than it might seem. While athletes with EIH often perform similarly to non-EIH athletes at sea level (suggesting compensatory adaptations), the picture changes at altitude. Durand and Raberin (2021) found that endurance performance during acute altitude exposure appears more impaired in EIH versus non-EIH athletes, because the combined effect of reduced ambient oxygen and an already compromised gas exchange system compounds the desaturation.
Altitude Training Implications
With the growing popularity of altitude training camps and live-high-train-low protocols, identifying which athletes have EIH becomes operationally important. An athlete who desaturates to 90% at sea-level maximal effort will desaturate further at 2,500 meters — and the performance consequences are more pronounced than for an athlete who maintains 96% at sea level.
Current identification methods require laboratory exercise testing with arterial blood gas measurement or high-quality pulse oximetry during incremental tests. A contactless, camera-based approach could eventually enable broader screening during routine training sessions, flagging athletes who warrant more detailed physiological testing.
Rehabilitation and Clinical Exercise
The application extends well beyond elite sport. Patients in cardiac rehabilitation, pulmonary rehabilitation, or post-surgical recovery programs need oxygen monitoring during prescribed exercise. Current practice often involves a nurse or therapist manually checking a pulse oximeter at intervals — a workflow that is labor-intensive and captures only snapshots rather than continuous data.
Camera-based monitoring in a gym or rehabilitation facility could provide continuous SpO2 trending for multiple patients simultaneously, alerting clinicians when any individual's oxygen saturation drops below a threshold during their exercise session.
Technical Approaches Under Development
The research community is pursuing several strategies to make camera-based SpO2 work during movement.
Multi-Region Adaptive Tracking
Rather than relying on a single facial region of interest, adaptive approaches track multiple zones and weight them by signal quality in real time. The forehead, for example, experiences less deformation during running than the cheeks. By dynamically re-weighting contributions from different regions, the system can maintain a usable signal even when some regions are temporarily corrupted by movement.
Deep Learning Motion Compensation
End-to-end neural networks trained on paired video and reference SpO2 data can learn to separate motion artifacts from physiological signals without explicit physics-based modeling. The challenge is training data — collecting synchronized facial video and gold-standard SpO2 during exercise is logistically demanding. Semi-supervised approaches like those proposed by Peng et al. (2024) reduce this burden by leveraging unlabeled video data.
Temporal Signal Stabilization
Filtering techniques adapted from inertial measurement and signal processing can stabilize the rPPG signal over time, using the known periodicity of the cardiac cycle to reject non-periodic motion noise. These work best during rhythmic activities like running or cycling where the motion pattern is somewhat predictable.
The Road From Here
Contactless SpO2 monitoring during exercise is not a solved problem. The honest assessment is that current camera-based systems work well at rest, show promise during light activity, and struggle with vigorous movement. But the trajectory of the research is encouraging, and the clinical demand is clear.
Circadify is developing camera-based SpO2 estimation technology designed to handle real-world conditions, including the movement and environmental variability that exercise introduces. The goal is not to replace clinical-grade pulse oximeters — those serve their purpose — but to bring oxygen monitoring into settings where clip-on sensors are impractical, uncomfortable, or simply absent.
Motion-robust rPPG algorithms are getting better. Deep learning is filling gaps that physics-based models leave open. And cameras keep getting cheaper and sharper. The athletes, patients, and clinicians who need this data are already waiting for it.
Frequently Asked Questions
Can a camera measure blood oxygen levels during exercise?
Emerging rPPG algorithms can estimate SpO2 from facial video during light-to-moderate activity, though accuracy degrades with intense motion. Researchers are developing motion-compensation techniques to extend reliability into higher-intensity exercise scenarios.
What is exercise-induced hypoxemia and why does SpO2 monitoring matter for athletes?
Exercise-induced hypoxemia (EIH) is a drop in arterial oxygen saturation during intense exercise, documented in endurance-trained athletes with SpO2 falling to 87-94%. Continuous SpO2 monitoring helps coaches and clinicians identify athletes affected by EIH and adjust training intensity accordingly.
How accurate is contactless SpO2 compared to a finger pulse oximeter during movement?
At rest or with minimal movement, camera-based SpO2 achieves errors under 4 percentage points. During exercise, motion artifacts increase error margins. Current research focuses on deep learning and multi-region-of-interest approaches to close the gap with contact-based devices.
Related Articles
- Contactless SpO2 Monitoring — A comprehensive look at how camera-based rPPG estimates blood oxygen saturation, including the underlying physics and clinical applications.
- Contactless Heart Rate Monitoring — Heart rate extraction forms the foundation of rPPG technology, and the motion challenges discussed here apply to heart rate as well.
- Environmental Factors in Contactless Vital Sign Accuracy — Lighting, skin tone, and ambient conditions all influence rPPG performance during exercise scenarios.