This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Biomedical Engineering, is properly cited. The complete bibliographic information, a link to the original publication on https://biomedeng.www.mybigtv.com/, as well as this copyright and license information must be included.
Many commodity pulse oximeters are insufficiently calibrated for patients with darker skin. We demonstrate a quantitative measurement of this disparity in peripheral blood oxygen saturation (SpO2) with a controlled experiment. To mitigate this, we present OptoBeat, an ultra–low-cost smartphone-based optical sensing system that captures SpO2我和心率而校准差异n skin tone. Our sensing system can be constructed from commodity components and 3D-printed clips for approximately US $1. In our experiments, we demonstrate the efficacy of the OptoBeat system, which can measure SpO2within 1% of the ground truth in levels as low as 75%.
The objective of this work is to test the following hypotheses and implement an ultra–low-cost smartphone adapter to measure SpO2: skin tone has a significant effect on pulse oximeter measurements (hypothesis 1), images of skin tone can be used to calibrate pulse oximeter error (hypothesis 2), and SpO2can be measured with a smartphone camera using the screen as a light source (hypothesis 3).
Synthetic skin with the same optical properties as human skin was used in ex vivo experiments. A skin tone scale was placed in images for calibration and ground truth. To achieve a wide range of SpO2for measurement, we reoxygenated sheep blood and pumped it through synthetic arteries. A custom optical system was connected from the smartphone screen (flashing red and blue) to the analyte and into the phone’s camera for measurement.
The 3 skin tones were accurately classified according to the Fitzpatrick scale as types 2, 3, and 5. Classification was performed using the Euclidean distance between the measured red, green, and blue values. Traditional pulse oximeter measurements (n=2000) showed significant differences between skin tones in both alternating current and direct current measurements using ANOVA (direct current:
In this work, we demonstrate that skin tone has a significant effect on SpO2measurements and the design and evaluation of OptoBeat. The ultra-low-cost OptoBeat system enables smartphones to classify skin tone for calibration, reliably measure SpO2as low as 75%, and normalize to avoid skin tone–based bias.
The measurement of blood oxygenation is necessary in acute medical emergencies and useful for tracking physical fitness [
Researchers have persistently documented how common pulse oximeters overestimate blood oxygen levels in patients with darker skin tones [
The COVID-19 pandemic has resulted in acute global shortages of necessary medical supplies, including pulse oximeters [
The proliferation of smartphones over the past decade provides a platform upon which a variety of health apps can be built. The combination of high-quality sensors, increasingly powerful mobile computation, and internet connectivity enables the creation of medical sensing systems that already live in billions of users’ pockets. More than 1 million mobile health apps are available on major mobile platforms [
In this paper, we introduce the equity-driven design of OptoBeat, a novel, ultra–low-cost smartphone-based pulse oximeter. OptoBeat can determine a coefficient to normalize pulse oximetry readings for differences in skin tone. OptoBeat can measure SpO2within 1% accuracy of the gold standard. We validated OptoBeat in blood oxygen levels that ranged from healthy (95%-100%) to critical (as low as 75%), corresponding to hypoxia. In this paper, we prove the following three hypotheses: (1) skin tone has a significant effect on pulse oximeter measurements (hypothesis 1), (2) images of skin tone can be used to calibrate for pulse oximeter error (hypothesis 2), and (3) SpO2can be measured with a smartphone camera using the screen as a light source (hypothesis 3).
Through quantitative analysis, we demonstrate how skin tone affects pulse oximeters. We show that the unilateral effect of skin tone not only decreases the signal to noise ratio but also affects the ratio of ratios between the 2 sources, which is used to calculate SpO2。We further demonstrate how this can be done with a smartphone and describe the design, development, and testing of an ultra–low-cost smartphone-based pulse oximeter. OptoBeat enables blood oxygen saturation monitoring by augmenting the smartphone’s camera system, focusing the light source, and leveraging extant computing capacity. With this system, we can use a skin tone measurement to adjust the ratios of the 2 source frequencies transmitted through the skin to calibrate the measurement of blood oxygen for a more accurate reading.
Specifically, we present the following contributions: (1) an experiment and quantitative analysis of how skin tone affects traditional pulse oximeters and a demonstration of how this can be remedied; (2) the design of the OptoBeat optical sensing system and an experiment to validate the theory behind our design; and (3) the design, execution, and results from an ex vivo experiment that validates the accuracy of pulse oximetry from healthy and critical SpO2levels against the gold standard and the results of a human-participant experiment to validate the efficacy of the device.
Typically, pulse oximeters measure the ratio of oxygenated to deoxygenated hemoglobin through capillaries in the finger, as shown in
Traditional pulse oximeter design.
Absorption spectrum of oxygenated and deoxygenated hemoglobin and common wavelengths used in pulse oximetry. Hb: hemoglobin; HbO2:氧ated hemoglobin.
Pulse oximetry measurements using standard commercial hardware can present biases and errors. Signal quality has been shown to decrease and increase with temperature when using finger-based measurements [
At saturation >80%, errors because of skin tone were not found to be significant; however, “in individuals with darkly pigmented skin, bias of up to 8% was observed at lower saturations” [
In our 20 yr [sic] of testing pulse oximeter accuracy, and probably in other testing laboratories, the majority of subjects have been light skinned. Most pulse oximeters have probably been calibrated using light-skinned individuals, with the assumption that skin pigment does not matter.[
This is similarly backed up by the 2013 Food and Drug Administration guidance on device calibration, which states that the “study should have subjects with a range of skin pigmentation, including at least 2 darkly pigmented subjects or 15% of your subject pool” [
Skin tone presents common complications in both dermal and transdermal optical measurements, particularly those in the visible spectrum. To help with this issue, researchers have found it useful to have a reference. One such standard reference is the Fitzpatrick skin tone scale. Although it is not without its own racial limitations [
Previous work has demonstrated how the Fitzpatrick scale was used to calibrate video-based heart rate monitoring in people with acute hypoxia [
The rapid development of smartphone sensors has enabled the design and implementation of smartphone-based medical and health-sensing systems. Systems that measure cardiac signals and blood composition, including heart rate [
Various apps have been developed in the past decade for mobile phone–based pulse oximeters [
Commercial smartphone apps lack detailed information on the underlying algorithms and analyses used [
Alternative smartphone-based systems incorporate an external sensor using the phone as a data hub. Many are standard fingertip-style pulse oximeters [
We validated the electromagnetic spectrum emitted by the screen (white, all LEDs at full brightness) using a spectrometer [
The OptoBeat system has several components to it, each of which we tested individually in a series of ex vivo experiments. These experiments allowed for control that was otherwise extremely difficult or impossible in a human-participant trial. For example, it would be extremely difficult to isolate skin from other components of the body (ie, blood, bone, and fat) in vivo. It is also dangerous for a healthy individual’s blood oxygen saturation to fall below 95%. To test the range of oxygen saturation required to satisfy the needs of the OptoBeat system, we had to control the full range of blood oxygen saturation required for medical diagnosis. In this section, we describe these experiments in detail as well as the design of the OptoBeat system and how it was used to test our three hypotheses (hypotheses 1-3).
The final design consists of three 3D-printed plastic clips, two 3-foot fiber-optic cables, 1 acrylic ball lens, and some rubber bands or O-rings for grip. The total cost is
Our goal was to focus as much light as possible and then couple the system with fiber-optic cables to maintain signal strength and block ambient light from entering the source (screen) or the receiver (camera). We evaluated dozens of commercial optical tools and systems to obtain an idea of what had been done; how we could move light around; and what materials were readily available, machinable, moldable, and affordable. This included optical filters, gratings, mirrors, and fiber plates, among others, that ranged from US $100 for individual components to tens of thousands of dollars.
Cross-sectional diagram of the OptoBeat system.
We limited our design exploration to materials and fabrication techniques such that the system could be built without specialized equipment. Most of the design, testing, and fabrication was executed using a laser cutter, a 3D printer, diamond files, and a drill press. In addition, to make an affordable, robust optical system, we explored a variety of raw materials. We found that the optimal shape for capturing light and coupling it with the fiber-optic cable was a long cone with a convex lens at the base. This design both collimated the beam and acted as a wave guide for coupling the light.
To make the lens, we experimented with hand-cutting a cast acrylic rod with diamond files and then polishing them to reach optical clarity. We also made a lens with a stereolithography 3D printer using clear resin that was then dipped in resin and cured to fill layer-height artifacts. In addition, we cast the negative of a lens (using a variety of materials) and then molded it with an optically clear urethane resin. Lens construction methods have proven to be labor-intensive, delicate tasks. Looking for alternative solutions to handmade or custom lenses, we found that a spherical lens worked nearly as well. This was mostly because focal length was not of concern as the distance to the fiber-optic cable could be easily adjusted. Cheap clear acrylic spheres are easily accessible from a variety of vendors and work well enough to focus light for transmissive pulse oximeters.
鉴于the limited brightness of the screen, we looked for other areas of the body in addition to the fingertip to capture SpO2。Our search focused on areas that were physically accessible, with a short optical path (distance between transmitter and receiver) and a high density of capillaries to improve the pulsatile signal. Three plausible solutions were the fingertip (same as traditional pulse oximeters), the webbing between the fingers, and the earlobe. Data were captured at each location, with the earlobe showing the highest signal to noise ratio for output magnitude and pulsatile signal.
The earlobe is comfortable, and a sensor could easily be worn for prolonged periods without impeding the user in most activities (no more than a set of wired headphones would), affording more continuous monitoring.
The OptoBeat signal acquisition app was designed on the iOS platform using Swift. The data acquisition page is shown in
同步相机屏幕刷新率和capture rate is crucial. The app ensures that the frequencies of data capture (camera) and source transmission (screen flashing) are synchronized. Misalignment results in the loss of essential information in the signal. An example of synchronization issues that we found is mixed-pulse capture. This occurs when half of the frame captures one color and the second half captures another. In our case, this was blue and red, resulting in a purple frame, which is unusable.
OptoBeat controls the camera using the Apple AVFoundation application programming interface. All automatic camera options are turned off, including autofocus, automatic white balance, and low light boost. The app allows the experimenter to control all other settings, including ISO, exposure time, and camera capture rate. The camera runs on a separate thread, and each frame of the video is sent immediately to the camera buffer. Data are directly accessed through the CMSampleBuffer (Apple Inc), and the RGB values are logged to the back end using a data management pipeline in a new thread without slowing down the main process.
The OptoBeat mobile app.
As the behavior of electromagnetic waves in different skin tones is not necessarily continuous, we decided to classify instead of regress across the spectrum of skin. To demonstrate hypothesis 1 and test hypothesis 2, we classified the different tones as 1-6 according to the Fitzpatrick scale. For a ground truth, we printed a copy of the scale to use as a reference and placed it beside the synthetic skin [
地面真理决心通过捕获mean RGB value of each reference color. For each synthetic skin sample, the mean RGB was calculated and iteratively compared with the 6 ground truth classes using Euclidean distance and the absolute difference in luminance-weighted grayscale (
The results were identical for both grayscale and RGB; an example image is shown in
Image of skin tone calibration strip and synthetic skin.
After validating that we had obtained an optical signal from the earlobe through the video recording analysis of our system, we designed an experiment to test hypothesis 3, validating OptoBeat’s SpO2measurements in mammalian blood using blue light pulse oximetry. In this experiment, we validated that, as hemoglobin deoxygenates, the transmission of blue light increases and that of red light decreases. This is based on the well-documented absorption and transmission spectrum of hemoglobin [
This setup demonstrated that the change in signal measurements could be attributed to blood oxygen saturation. After being oxygenated as in the previous experiment, sheep blood was pulsed through the artery at 60 beats per minute (bpm) to mimic a heart rate, and the blood was fed back into an open beaker, where it was exposed to standard air pressure and the hemoglobin could continue to deoxygenate.
Cross-sectional diagram of the blood oxygen saturation experimental setup.
To validate that the OptoBeat system would work on human participants, data were collected on three of the authors. Data were collected 3 times for each author. The data collection period was approximately 30 seconds for each sample.
As there was no risk and the participants were all authors of this paper, we did not require Institutional Review Board approval. Data were collected between November 2020 and November 2021.
The first experiments described in this section demonstrate the effect of skin tone on pulse oximeters. All the measurements in this first experiment are direct current (DC). These DC-component experiments demonstrate how these attributes affect the signal quality when isolated from the much more complex human body.
We measured the transmission of red and IR light in the 3 different skin tones described in the
By accessing the photodiode output of a commodity pulse oximeter, the values of the 660-nm (red) and 880-nm (IR) light as it passed through the different samples were recorded. It was important to record these values and not the resulting SpO2values as they had already been fitted to the model. To show that the problem lay in the actual hardware, we compared the red and IR values (
For each skin tone, 2000 samples were collected, and the ratio of IR to red was calculated (type 5: SD 0.42%; type 3: SD 0.25%; type 2: SD 0.25%). We ran an ANOVA test and a pairwise comparison of the results. The results showed that the ratio of transmission for each skin tone was significantly different from the others (
To demonstrate hypothesis 2 (the use of skin tone data in calibrating pulse oximeters), the median sample, classified as type 3, was used as a reference to calibrate the others. For each skin tone, we derived coefficients to normalize the varying ratios between the different skin tones.
The ratio of the ratios was used to produce the following coefficients to normalize the skin tone absorbance of red and IR. The experiment was then replicated with our OptoBeat system using blue and red light, as shown in
If it were the case that both wavelengths of light were affected equally, then the only errors that would arise would be due to changes in the signal to noise ratio. However, the change was unilateral, affecting the ratio of the wavelengths and not just the signal strength. Using the ratio of ratios between the alternating current (AC) and DC components of each wavelength did mitigate this to some extent. However, we hypothesized that the AC to DC ratio would also gain error across skin tones owing to their vastly different absorbency characteristics. To confirm this and further validate hypothesis 1, the same experiment as above was conducted using a pump system to move distilled water through a synthetic artery [
As shown in
Using an ANOVA test, the results were statistically significant (
Results from a direct current pulse oximeter skin tone experiment.
Skin tone type | Infrared value | Red value | Ratio of ratios | Difference from type 5, % | Difference from type 3, % | Difference from type 2, % |
5 | 0.49 | 0.86 | 0.58 | N/Aa | 12.1 | 14.8 |
3 | 0.51 | 1 | 0.51 | 12.1 | N/A | 3.2 |
2 | 0.50 | 0.96 | 0.52 | 14.8 | 3.2 | N/A |
aN/A: not applicable (data are compared to themselves).
Results from a direct current OptoBeat skin tone experiment.
Skin tone type | Blue value | Red value | Ratio of ratios | Difference from type 5, % | Difference from type 3, % | Difference from type 2, % |
5 | 0.8 | 0.61 | 0.76 | N/Aa | 19.3 | 16.6 |
3 | 1 | 0.94 | 0.94 | 19.3 | N/A | 3.3 |
2 | 0.65 | 0.59 | 0.91 | 16.6 | 3.3 | N/A |
aN/A: not applicable (data are compared to themselves).
Results from an alternating current pulse oximeter skin tone experiment.
Skin tone type | Ratio of ratios | Difference from type 5, % | Difference from type 3, % | Difference from type 2, % |
5 | 0.77 | N/Aa | 4.3 | 17.7 |
3 | 0.81 | 4.3 | N/A | 14 |
2 | 0.94 | 17.7 | 14 | N/A |
aN/A: not applicable (data are compared to themselves).
The data (n=400; 10-second samples, 67 minutes total) were cleaned with a bandpass filter, with cutoff frequencies at 0.5 Hz and 2 Hz (30-120 bpm pass band), and then passed through a Savitzky-Golay filter for further smoothing (third order, 35-sample window). This provided a clear pulsatile signal, as shown in
The resulting
OptoBeat measures continuously with a floating point, whereas the ground truth estimates integer values. We believe that the performance would be more strongly correlated if both were continuous. We plan to evaluate this in the future.
Pulsatile signal of red and blue lights captured in the oxygen saturation experiment.
Results of the quadratic support vector machine (SVM) mapping R values captured from OptoBeat to the ground truth. SpO2: peripheral blood oxygen saturation.
Predicted versus true response from regression.
Residual error: OptoBeat and the ground truth. SpO2: peripheral blood oxygen saturation; SVM: support vector machine.
Each of the participants (N=3; samples=3 × N, duration=20-30 seconds/sample) measured within –1% to +1% of the ground truth, a commodity pulse oximeter.
Pulsatile signal of red and blue lights captured in the human-participant proof-of-concept experiment.
OptoBeat’s peripheral blood oxygen saturation calculations against the ground truth in the human-participant proof-of-concept experiment.
In this paper, we have presented the design and evaluation of OptoBeat, an optical attachment for smartphones that can reliably measure SpO2and calibrate the measurements according to skin tone via images. The phone-based system for oxygen saturation measurement has potential benefits compared with existing pulse oximeters. With the added benefit of the smartphone’s hardware and computing power, we can not only measure SpO2but also skin tone, which can be used to account for errors in SpO2measurements.
Our system is cheaper and simpler to produce than most commercial pulse oximeters. Manufactured at scale, it could be used to shift from a regime of rigid thresholds for admittance or treatment at the point of care to home measurement. At home, trending measures of pulse oximetry could be used to monitor clinical progress or detect silent hypoxemia early [
Although this work demonstrates the efficacy of OptoBeat in an ex vivo laboratory experiment, our human-participant proof of concept was only meant to show feasibility, not to validate it for human use. Clinical studies would have to be conducted on a large, diverse population to validate this.
前进,我们计划部署和测试的坏蛋rent OptoBeat system in a clinical setting. We have partnered with our medical school to take measurements while patients are undergoing cardiothoracic surgery as this will give us access to a range of blood oxygen measurements while a ground truth is being collected without adding additional risk to participants. Furthermore, we plan to redesign traditional pulse oximeters using what we learned from OptoBeat to develop a stand-alone device that can account for variations in skin tone. This will include a large-scale skin tone data collection to build a model that can be leveraged by the device.
可穿戴设备目前站加剧现有sting health disparities in underserved racial populations who would benefit most from enhanced detection and treatment of health issues. Indeed, such racial disparities occur in the context of more significant cultural issues and reflect the mistrust that many underserved racial populations have for the medical system. This mistrust is primarily due to the medical community’s historical bias toward addressing White Americans’ health needs at the expense of underserved racial populations’ health and well-being. There is an urgent need to overcome the cycle of disparities produced by social injustice across conditions in the places where people live, learn, work, and play. Ensuring that wearables and remote patient monitoring tools are equally efficacious across different populations is necessary to accelerate health equity in the populations who would benefit most from such technology.
This project was supported by division of Information and Intelligent Systems (RAPID; award number 2031977) of the National Science Foundation. It was also supported by the National Institute on Drug Abuse (grant K23DA041616), the National Institute on Minority Health and Health Disparities (grant P50MD017347), and the National Institute of Allergy and Infectious Diseases (grant P30AI110527) of the National Institutes of Health.
Skin tone calibration equations.
alternating current
beats per minute
direct current
infrared
light-emitting diode
red, green, and blue
peripheral blood oxygen saturation
TC is the cofounder of HealthRhythms, Inc, and also the Senior Vice President of Digital Health at the Optum Labs (part of United Health Group). The other authors have no conflicts to declare.