Adaptive Pulse Oximeter Improves Oxygen Measurement Across Skin Tones

This can contribute to occult hypoxemia, in which dangerously low arterial oxygen levels go undetected by routine readings, a problem reported more frequently in Black patients. Because accurate oxygen measurements are critical for decisions about oxygen therapy, hospital admission, and intensive care, researchers at Tufts University have developed a wrist-worn pulse oximeter that automatically adapts to a patient’s skin tone.

The device, called ChromaSense, measures blood oxygen saturation, heart rate, and respiration rate using a watch-sized module worn on the wrist. Rather than relying on light transmitted through the finger, it analyzes light reflected from the skin and underlying tissue. Before monitoring begins, the device captures each user’s skin reflectance profile and adjusts both emitted light intensity and signal-processing parameters, allowing measurements to be tailored to individual skin tone.

ChromaSense is based on photoplethysmography, which measures pulsatile changes in microvascular blood volume. Oxygen saturation is calculated from the relative absorption of red and infrared wavelengths during the arterial pulse. Because melanin also absorbs and scatters light, darker skin can weaken the optical signal or alter the ratio used to calculate saturation. The system’s individualized calibration is designed to reduce these errors at both the measurement and algorithmic levels.

Concerns about bias in conventional pulse oximeters gained prominence during the COVID-19 pandemic, when a 2020 New England Journal of Medicine study reported occult hypoxemia in up to 17% of Black patients, more than three times the rate observed in white patients. In early testing involving 50 participants across a range of skin tones, ChromaSense measured oxygen saturation within 1.4% of a standard reference oximeter at normal saturation levels. In a new study at the Hypoxia Research Laboratory at the University of California, San Francisco, healthy volunteers were briefly desaturated across a 70% to 100% range. The device achieved accuracy within 2.87% of a blood-referenced oximeter, meeting U.S. Food and Drug Administration (FDA) performance requirements without observable skin tone-dependent bias, including at lower oxygen levels.

The researchers are also exploring cuffless blood pressure monitoring using photoplethysmography (PPG) waveforms. Machine-learning models trained on large healthcare databases were evaluated in 2,315 adult intensive care unit patients across race, gender, age, and combined subgroups, achieving reported accuracy of up to 90% for systolic and diastolic blood pressure. This capability has not yet been integrated into ChromaSense.

“The blood-pressure work is not yet built into ChromaSense, but the goal is in the future to embed that machine learning model into the device,” said Valencia Koomson, associate professor of electrical and computer engineering at Tufts University.

“If you train a model that converts light signals to blood oxygen, pulse or pressure and you don’t ensure that the dataset that you’re training with is diverse enough in terms of age, race, and gender, it can affect the performance or accuracy of the model. An apparently high-performing model can look far less impressive once broken down into specific groups,” said Koomson.

Related Links
Tufts University