If two people have the same oxygen saturation, should the same fingertip sensor give them the same reading? A pulse oximeter estimates blood oxygen by shining red and near-infrared light through a finger and comparing how much each wavelength is absorbed. That is a physical measurement, but the finger is not an empty tube. Skin pigment, blood flow, movement, and device calibration all influence the light reaching the detector—and can shift the estimate. [1]
A reading is built from changing light
Oxygenated and deoxygenated blood absorb red and near-infrared light differently. A pulse oximeter uses those differences to calculate an estimate of oxygen saturation. It does not directly count oxygen molecules in a blood sample; it interprets light after that light has passed through tissue. [1]
That distinction matters because the device’s signal has more than one source. The blood contributes absorption, but skin and other tissue also interact with the emitted wavelengths. The oximeter must use a pattern in the changing signal to estimate the blood-related component. Think of trying to hear one instrument in a room where several people are talking: a stronger background can make the target harder to separate even when the instrument itself has not changed.
The same interference need not be equal
NHLBI describes several situations that can reduce pulse oximeter accuracy: low oxygen saturation, poor fingertip blood flow, nail polish, and physical motion. Blood loss or constricted vessels can also impair the signal. Motion artifact—the disturbance produced when movement changes the measured signal—can be especially relevant when a person’s oxygen level changes during activity. [1]
Skin pigment is another source of unequal interference. Melanin absorbs across a range of wavelengths, including the red and near-infrared light used by many pulse oximeters. NHLBI reports that readings tend to overestimate oxygen saturation in patients with darker skin tones. If an estimate is biased upward, a low true value may appear less concerning than it is. The error is not merely a technical footnote when a measurement informs decisions about care. [1]
An algorithm inherits the limits of its examples
Artificial intelligence can identify patterns in sensor data, but it learns from examples assembled into datasets. For a medical imaging algorithm, NIBIB notes, records need curation, quality assessment, annotation, and expert labeling. Researchers also need to check that tools perform fairly and equally across groups and populations. A model trained on examples that poorly represent some people may have less evidence from which to learn their signals. [2]
This is related to, but distinct from, a sensor’s optical bias. A pulse oximeter can produce a distorted input because of how light interacts with skin; an algorithm can add another layer of error when it interprets patterns using unevenly representative data. More computation does not automatically remove a physical limitation. Nor does a large dataset guarantee a fair model if its examples or labels fail to capture relevant differences.
Better measurement means testing the whole chain
Researchers are exploring pulse oximeter designs that use green light, which has a larger absorption difference between oxygenated and deoxygenated blood. The described approach measures reflected light rather than transmitting it through the finger, and its developers calibrated it using a skin-pigment color scale. Those are design strategies under study; they do not establish that every new device will solve every accuracy problem. [1]
A useful way to evaluate a health measurement is to follow the full chain: what physical signal is emitted, what tissues alter it, what detector records, and what calculation turns that record into a number. AI systems add another link: which examples taught the model what patterns mean. Testing across populations and conditions can reveal whether a device’s errors cluster in particular groups. Fairness begins with that concrete question about the instrument’s interaction with real bodies. [2]
Sources & further reading
- Updating Pulse Oximeters NHLBI
- Artificial Intelligence NIBIB
Prepared for the independent PSI Center reading library. About edition dates and source checking.
This explainer describes scientific principles. It does not assess individual treatment choices.
