October 14, 2026 | Melanin absorption has been a recognized if widely overlooked factor in pulse oximetry for decades, which can result in a measurement bias that overestimates oxygen levels in people with darker skin tones. A device under development in the Advanced Integrated Circuits and Systems Lab at Tufts University seeks to force attention to the issue while also adding a blood pressure monitoring feature powered by machine learning, according to the lab’s founding director Valencia Koomson, Ph.D., associate professor of electrical and computer engineering.
Melanin is a “chromophore,” a naturally occurring substance in the body that absorbs and scatters light and has a profound effect on pulse oximeter readings, which was dramatically highlighted during the COVID-19 pandemic, says Koomson. A retrospective study led by a critical care physician in Michigan found that dangerously low oxygen went undetected more than three times as often in Black patients as in White patients (New England Journal of Medicine, DOI: 10.1056/NEJMc2029240)
As a result, the U.S. Food and Drug Administration (FDA) posted a formal safety communication regarding the risk of inaccuracy in pulse oximeters. But interviews conducted in 2024 with more than 100 physicians at local hospitals in the Boston and Baltimore area suggest that less than 30% of doctors are aware of this FDA safety advisory, says Koomson, and many of them don’t even fully understand how the devices work.
A series of small-scale studies conducted at the Hypoxia Lab at the University of California, San Francisco, over the past couple of decades have all found that off-the-shelf pulse oximeters read oxygen levels too high in people with darker skin when oxygen levels are in fact low, she says. But, disturbingly, textbook descriptions of near-infrared and light-based sensors for medical monitoring rarely mention melanin absorption.
This put Koomson on a mission to develop ChromoSense, a wrist-worn wearable that can measure blood oxygen saturation, heart rate, and respiration rate across a broad range of skin tones. The chief novelty of the device is that it first analyzes an individual’s specific skin reflectance profile to account for different levels of melanin before auto-adjusting its light emission and signal processing, as described in a research article that was published recently in Review of Scientific Instruments (DOI: 10.1063/5.0304023).
The reflectance-based sensor eliminates the historical inaccuracies and skin-tone biases found in standard oximeters, which generally measure light transmitted through the finger. In the latest study, this was demonstrated through a dynamic, hardware-level calibration process rather than static software assumptions, she explains. In Hypoxia Lab testing, where oxygen levels were deliberately lowered across a 70% to 100% saturation range in healthy volunteers, ChromoSense also performed well against a reference oximeter that reads directly from the blood.
That testing was done for free, thanks to a grant from the Hypoxia Lab, which funds research teams working on solutions to pulse oximeter limitations, Koomson reports. This negated the need for her and her team to recruit study participants and hire an anesthesiologist and medical staff to monitor their safety.
Koomson joined Tufts to establish the Advanced Integrated Circuits and Systems Lab in 2005, the initial focus being understanding how semiconductor technology and microelectronics could be used to improve health outcomes and improve overall quality of life, she says. One of her first collaborations was with researchers in the bioengineering department working on diffuse optical imaging using near-infrared light, a primary driver behind the miniaturization and portability of medical imaging devices.
The overarching goal has been to leverage modern computing technology to improve the performance of these devices using near-infrared optical signals to noninvasively measure different biomarkers in the body, including blood flow and oxygen in tissue, continues Koomson. Among the many advantages of using near-infrared signaling for medical monitoring is that the systems can be built using low-cost, off-the-shelf components and integrated along with microchips and nanoscale circuit systems to enable a range of diagnostic capabilities that aren’t practical to do using large, bulky instruments.
A grant from the National Science Foundation is providing support for miniaturized, near-infrared systems that can capture hemodynamic response within brain tissue, notably to look at how oxygen is flowing to the brain of preterm infants who might otherwise have to be placed into an MRI machine. Inadequate oxygen flow in this vulnerable population could have long-term neurological impacts, she points out.
The pivot to pulse oximetry was prompted by the pandemic, and the study out of Michigan about how the fingertip devices tended to report a falsely high oxygen percentage in people with darker skin tones, she shares. These standard pulse oximeters use near-infrared light as part of a dual-wavelength system to measure blood oxygen saturation.
Pulse oximetry is well-established technology, but the devices haven’t been designed in a way that accounts for all the various biological media that can absorb and scatter light and can therefore throw off readings, says Koomson. The FDA standard for marketing clearance is testing in at least 10 people, including males and females of different ages and at least 15% of the test population having dark skin pigmentation. The problem is that dark skin pigmentation is “very relative ... [and] not well defined,” she adds, and the question has come up in forums over the past few years about whether the sample size is adequate to characterize a device that’s used on such a diverse population of people.
This is a worldwide issue, so the affected population is huge. And melanin absorption and scattering are dominant factors, but not the only ones affecting the accuracy of results, Koomson says. So are low blood perfusion levels, current smoking status, and fingernail polish if the sensor is being placed on the finger.
But in comparing the accuracy of pulse oximeters in Black versus White patients, skin pigmentation is the only relevant confounding factor. In the pandemic-era retrospective study, looking back at data in electronic health records, it was the only thing that could explain the discrepancy in results when pulse oximeter readings were compared to the gold-standard blood gas reading, she points out.
In addition to interviews with physicians, Koomson says she has spoken to medical device manufacturers and presidents of patient advocacy groups in respiratory health about the uneven performance of pulse oximeters. She found that physicians are often unwilling to “accept the fact that a device they have been relying on so heavily is flawed.”
The FDA, meanwhile, holds that pulse oximeters are not meant to be diagnostic devices but to capture trends, she adds. Agency representatives “argue that a lot of physicians are using pulse oximeters in the way they should not be used.”
Relative to arterial blood gas testing, pulse oximetry is faster, painless, and doesn’t require a trained nurse to do the measurement—and this is why Koomson is looking to improve rather than replace the technology. Typically, attempts to correct accuracy issues in modern medical devices involve software fixes to minimize or eliminate errors once a signal has been captured, she says.
Koomson’s approach is to instead understand the sources of noise that the light signal in the sensor is going to encounter and try to optimize how much signal passes through the skin. Doing that requires assessing the level of melanin absorption and scattering that’s going on in the tissue, she says in describing the hardware of ChromoSense.
The reflectance-based device is designed to enhance the signal-to-noise ratio and give users feedback when the signal is too low to provide an accurate reading. “Most pulse oximeters on the market today don’t do that,” says Koomson. “They give a default reading, even if the patient might be deceased.”
Medical errors, including those related to not having an accuracy indicator on pulse oximeters, get reported via the FDA’s MedWatch program. Koomson has also heard from nurses doing home-based monitoring about patients who are clearly in respiratory distress but getting healthy-level readings, and how they futilely tried to fix the problem by changing the batteries or using a different finger.
While most pulse oximeters only use two wavelengths of light, ChromoSense utilizes additional wavelengths of light to eliminate skin tone bias and melanin-induced optical attenuation, she explains. This also corrects for age-related biases related to the skin getting thinner, which with conventional pulse oximeters results in less of the light traversing through a body part—be it finger (most commonly), ear, nose, or anywhere else—from being absorbed and scattered. A construction worker’s thick and calloused fingers can likewise introduce errors by physically changing how light travels through tissue when off-the-shelf devices are used, she says.
“Pulse oximetry is not a direct measure of blood oxygen; it’s based on an empirical formula,” she notes. “We measure light at different oxygen levels and try to fit a line to the data points to say if the sensor gets this reading, to correspond that to an oxygen saturation level.”
Koomson and her team are soon to publish another paper expanding on results they presented at the International Conference of the IEEE Engineering in Medicine and Biology Society in July related to characterization of the light output (light-emitting diode) of ChromoSense. They have estimated the spectral shift that can be expected from the device over time and could help explain some of the errors seen in some pulse oximeter studies.
The test equipment is available at Tufts, she says. The analysis will additionally look at how electrical components of the device perform over different temperatures.
“The next exciting thing we want to add is blood pressure monitoring,” says Koomson, who has had a Ph.D. computer science student working in her lab for the past two years on blood pressure estimation using the photoplethysmogram (PPG) waveform being captured by the pulse oximeter. “Most people just use it [PPG] to extract oxygen saturation, but I feel like it is a very much an under-studied biometric waveform.”
Machine learning models have already been used to parse out different subgroups by skin tone, age, and gender when estimating systolic and diastolic blood pressure from PPG waveforms found in large healthcare databases, using the width of different parts of the waveform to indicate the level of stiffness in arteries.
Changes in vascularization in the body due to individual physiology and age—e.g., the amount of constriction in the arteries as blood is pumped from the heart to the extremities—can directly affect the shape of the PPG waveform that’s captured by the sensor, as can blood pressure, she says. “A number of research groups around the world are trying to design devices that can estimate blood pressure by extracting features from these waveforms ... but what’s new is we’re trying to use machine learning inference models to do that.”
A paper is soon to be published introducing a way of ensuring different subgroups are represented in training datasets for machine learning models making blood pressure predictions, says Koomson. Historically, some published articles on blood pressure estimation models have reported excellent accuracy, but with subgroup analysis performance dips at the intersection of race, gender, and age—e.g., Asian men between 18 and 25 or White women over 65.
Kofimi Technology, Inc., is an early-stage medical device company that spun out of Koomson’s lab. The startup is using support from Small Business Innovation Research grants and non-diluted funding to build a prototype and do more testing of the ChromoSense device in different clinical settings. A large-scale study enrolling patients with respiratory or lung diseases is now being planned.