The LED Is Part of the Problem, and That Part Already Has a Fix

A pulse oximeter works by shining red and infrared light through a finger and reading the ratio of what comes back. That ratio maps onto oxygen saturation through a calibration curve, and the whole device is only as good as the assumption that the curve is the same for everybody.

In 2026 two physicists modeled what happens to that curve when you change one component. Not the algorithm. The light source.

With the broad-linewidth light-emitting diodes that commercial oximeters use, the calibration curve depends on how much melanin is in the epidermis. With monochromatic laser diodes, it does not.

Their conclusion is that spectral coloring from the LEDs contributes to the skin-tone bias, and that adopting laser diodes could eliminate that contribution altogether.

An LED does not emit one wavelength. It emits a band. When that band passes through a layer that absorbs the short end more than the long end, the effective wavelength that arrives is not the wavelength the calibration assumed — and the device is now solving the wrong equation, slightly, in a direction that depends on the patient.

How big the clinical version of this is

A 2024 systematic review and meta-analysis pooled 15 studies covering 732,505 paired oximetry measurements from 207,464 hospitalized patients.

Occult hypoxemia — the oximeter reading reassuringly while the arterial blood is genuinely low — was more common in Black patients, pooled prevalence ratio 1.67. For patients identifying as Asian, Latinx, Indigenous, multiracial or other, 1.39.

A 2026 study in the BMJ measured it prospectively rather than from records. Twenty-four English intensive care units, 903 critically ill adults, 11,018 paired readings against arterial blood, and — this is the part that matters methodologically — skin tone measured objectively with a handheld spectrophotometer rather than inferred from race.

Every oximeter tested overestimated at low saturations and underestimated at high ones. On average, readings ran 0.6 to 1.5 percentage points higher for patients at the dark end of the individual typology angle scale.

One percentage point sounds trivial. It is not trivial at 92, which is the number that decides whether somebody gets escalated.

The evidence that complicates it, which I am not going to bury

This finding has become a headline, and headlines flatten things.

A prospective neonatal study published in 2026 enrolled newborns with arterial lines in two intensive care units, took 136 matched pairs from 70 patients, 40 percent of them Black, and measured pigmentation three ways — individual typology angle, melanin index, and visual assessment.

Across the categorical measures and across race, it found no statistically significant difference in bias. Analyzed continuously, the bias did shift with pigmentation, but the effect reached significance only for one of the three measures and only when the analysis was restricted to the first reading per patient.

That is a well-designed study, in the population where getting it wrong matters most, and it mostly did not find the effect.

The bench work is equivocal too. A laboratory group built a test rig with synthetic melanin filters and ran six commercial devices through simulated saturations. The devices disagreed wildly with each other, especially below 80 percent — but the melanin filters did not change their responses, and the authors say plainly that their bench tests do not provide conclusive evidence either way about what melanin does in a living finger.

And the physics has a caveat with teeth. A Monte Carlo simulation of a four-layer finger found that in transmission mode — light passing straight through, which is what a fingertip clip does — the calibration curves did not deviate with skin tone at all, because both the pulsing and non-pulsing components were attenuated equally at both wavelengths. The deviation appeared in reflectance mode, where source and detector sit on the same side, and it shrank as the two were moved further apart.

So the honest summary is not oximeters are broken on dark skin. It is that a small, real, direction-consistent bias exists in the large clinical datasets, that it concentrates at low saturations where the measurement matters most, and that the mechanism is a combination of sensor geometry and light-source physics rather than a single simple flaw.

What the fix looks like

A pilot study took 18 people, nine darkly pigmented and nine lightly pigmented, and compared two conventional LED oximeters against a laser-based one, all against arterial blood.

The laser device had significantly lower measurement error — variance of 3.54 against 5.48 and 5.47, at p below 0.001.

Eighteen people is a pilot and I would not call it settled. But it is the same answer the simulation gave, arrived at independently, and it points at something unusual in this section: a bias with an engineering solution that already exists.

Most of the failures documented across these pages require somebody to go back and run the study nobody ran. This one requires a component substitution.

The same physics, in a different room

Functional near-infrared spectroscopy shines light into the scalp to watch blood oxygenation in the brain. It is popular in developmental neuroscience precisely because it tolerates a wriggling child better than a scanner does.

And participants with Afro-textured hair have routinely been excluded from it, because the sensors cannot make contact with the scalp.

Not excluded by policy. Excluded by the equipment not working, and then by researchers quietly not recruiting people it does not work on — which produces a dataset that looks complete and is not.

A 2026 paper set out to fix the practical end of this. The team built an intake survey, staffing and training guidance, and a set of hair preparation techniques: braiding cornrows around the intended sensor layout, using gels and clips to part the hair, and various ways of increasing tension to get the optode onto skin. They tested it with 19 Black participants, eleven adults and eight children, across a range of hair types and styles.

Signal quality improved by an average of 50 percent, most in the frontal regions. Not perfect — the authors say so — but good enough for those participants to be included in a dataset rather than screened out of one.

I find this the most quietly damning study in this article, because of what it implies about the years before it. Every fNIRS finding published while this was unsolved was published on a sample from which a specific group had been removed by a hardware limitation, and almost none of those papers say so.

The trend, the symptom, and the blood gas

If you use a home oximeter, read the trend rather than the digit. The bias in these studies is roughly one percentage point on average and larger at the bottom of the range, so a single reading of 93 is not the same information as a reading of 93 in somebody else, but a fall from 97 to 93 over two days is your reading falling on your own device.

Symptoms outrank the number. Breathlessness at rest, confusion, difficulty finishing a sentence — none of those becomes less real because a device says 94.

In hospital, the question that resolves it is an arterial blood gas, which measures oxygen in the blood directly rather than inferring it through skin. That is the gold standard every one of these studies compared against, and asking whether it has been checked is a reasonable question rather than a challenging one.

And if you own a device, transmission-mode fingertip clips are what the simulation found least affected. That is a modeling result rather than a purchasing recommendation, but between a fingertip clip and a reflectance sensor worn on a wrist, the physics currently favors the finger.

This article is educational and is not a substitute for care from someone who can look at your skin. Anything new, changing, painful or spreading should be assessed rather than researched.