⌚ Wearable SpO2 Pulse Oximetry Skin Tone Bias

Two-Wavelength LEDs — the Physical Basis of Pulse Oximetry

Every finger-clip or wrist-worn pulse oximeter relies on a strikingly simple idea from 1970s Japanese biomedical engineering (Aoyagi, 1974): shine two colors of light through living tissue and watch how the color balance changes with each heartbeat. Oxygenated hemoglobin (HbO2) and deoxygenated hemoglobin (Hb) absorb red and infrared light differently, and that difference — measured continuously — is converted into a percentage oxygen saturation reading updated every second.

• 660 nm: Red LED wavelength (strongly absorbed by deoxy-Hb)
• 940 nm: Infrared LED wavelength (strongly absorbed by oxy-Hb)
• 250 Hz: Sampling rate (typical) (per wavelength channel)
• 1974: First clinical device (Aoyagi, Nihon Kohden)

Why red and infrared light

Hemoglobin's absorption spectrum crosses over near 805nm — the "isosbestic point" where oxy- and deoxy-hemoglobin absorb light equally. Away from that point, the two forms diverge sharply:

• At 660nm (red): deoxygenated hemoglobin absorbs roughly 10x more light than oxygenated hemoglobin • At 940nm (infrared): the relationship reverses — oxygenated hemoglobin absorbs somewhat more than deoxygenated

By comparing how much red vs. infrared light makes it through the tissue, the device infers the ratio of oxy- to deoxy-hemoglobin in the arterial blood, without ever drawing a drop of blood.

Wearables (rings, watches) typically use reflectance mode — LEDs and photodetector sit side-by-side on the same surface, measuring light that bounces back through skin — rather than the transmission mode used by fingertip clips. Reflectance geometry is more sensitive to skin surface properties, including melanin content, because light must pass through the epidermis twice.

The entire measurement pipeline assumes that the only thing changing the red/infrared color balance with each heartbeat is arterial blood oxygenation. Any other light-absorbing structure in the optical path — including melanin — introduces a systematic error the algorithm cannot distinguish from a true physiological signal.

Anatomy of the optical path

Light emitted by the LEDs does not travel in a straight line to the detector. It scatters through multiple tissue layers, each with different optical properties:

• Stratum corneum & epidermis (0.05–1.5mm): contains melanocytes and melanin granules; largely avascular • Dermis (1–4mm): contains capillaries, small arterioles, and venules — the pulsatile blood signal originates here and in the deeper subcutaneous vessels • Subcutaneous fat and connective tissue: mostly static absorbers, contribute to the DC baseline • Nail bed / bone (fingertip devices): additional static attenuation

Melanin sits in the very first layer light must cross — before the light ever reaches the blood vessels the sensor is trying to interrogate, and again on the way back to the detector in reflectance-mode wearables.

Melanin Absorption and Scattering — the Beer-Lambert Law in Real Skin

The Beer-Lambert law describes how light intensity decays exponentially as it passes through an absorbing medium: I = I0 × e^(−ε·c·l), where ε is the wavelength-specific absorption coefficient, c is the concentration of the absorber, and l is the path length. Pulse oximetry algorithms apply a version of this law assuming hemoglobin is the only variable absorber. Melanin violates that assumption.

• 1–43%: Melanin content range (epidermal volume fraction, type I→VI)
• ~1.5x: Melanin absorption 660nm (higher than at 940nm)
• up to 10x: Light attenuation, dark skin (more signal loss vs. fair skin)
• ↓ SNR: Effect on signal-to-noise (more amplifier gain, more noise)

Melanin as a confounding chromophore

Melanin (predominantly eumelanin in darker skin) is a broadband absorber — it absorbs light across the visible and near-infrared spectrum, but not uniformly. Its absorption coefficient decreases roughly monotonically with wavelength, meaning it absorbs somewhat more red (660nm) light than infrared (940nm) light, though both are attenuated substantially compared to unpigmented skin.

Because pulse oximetry's entire signal depends on the relative red vs. infrared attenuation, any wavelength-dependent absorber other than hemoglobin distorts the ratio the algorithm depends on. Melanin's uneven spectral fingerprint means it does not cancel out — it systematically skews the red/infrared balance in a way that correlates with pigmentation level, not with true oxygenation.

Epidermal melanin content varies enormously by Fitzpatrick skin type: from roughly 1–3% volume fraction in Type I skin to over 40% in Type VI skin. This is a continuous physical variable, not a simple binary — which is why bias scales gradually with pigmentation rather than appearing only in the darkest skin.

Beer-Lambert law applied to layered, scattering skin

Real skin is not a simple absorbing solution — it is a highly scattering, layered medium, so the true light transport is described by the more complex diffusion approximation of radiative transfer, not pure Beer-Lambert. But the intuition holds:

I(λ) = I0(λ) × exp[−(ε_Hb(λ)·c_Hb + ε_mel(λ)·c_mel + μs'(λ))·l]

Where ε_Hb·c_Hb is the hemoglobin absorption term the device wants to measure, ε_mel·c_mel is the melanin absorption term acting as noise, and μs' is the scattering coefficient (also elevated by melanin granules and by higher collagen density in some skin types).

Because melanin sits in a fixed anatomical layer, it contributes mostly to the DC (non-pulsatile) component of the signal — but it also reduces the total light reaching the detector, forcing automatic gain control to boost amplification. This raises electronic noise relative to the tiny AC pulsatile signal, degrading the signal-to-noise ratio the ratio-of-ratios calculation depends on.

Why wearables are more exposed than hospital fingertip clips

Reflectance-mode wearables (rings, watches) face compounded challenges relative to hospital-grade transmission fingertip probes:

• Shorter, shallower optical path: light must scatter back to a detector on the same side, increasing the fraction of the path spent in melanin-rich epidermis relative to blood-rich dermis • Motion artifact: wrist-worn devices see more relative motion between sensor and skin than a clipped fingertip, compounding signal loss • Variable contact pressure and perfusion: looser fit reduces capillary blood volume signal (perfusion index), which is already weaker in low-perfusion or cold extremities • Consumer-grade LEDs and photodiodes: typically lower dynamic range than clinical pulse oximeters, offering less margin to compensate for melanin-induced signal loss

All of these effects compound with pigmentation, meaning wearable-specific bias can, in principle, be larger than what has been documented for clinical fingertip devices.

The Photoplethysmogram — Extracting a Heartbeat From a Sea of Static Light

The raw signal reaching the photodetector is dominated by light that never touched arterial blood — skin, bone, venous blood, melanin. Only a tiny fraction, roughly 1-2% of total transmitted light, varies with each heartbeat as arterioles fill and empty. This pulsatile component, the photoplethysmogram (PPG), is the signal pulse oximetry must isolate and amplify.

• 1–2%: AC pulsatile fraction (of total detected light)
• 98–99%: DC baseline fraction (skin, bone, venous blood, melanin)
• 0.2–20%: Perfusion index (typical) (AC/DC ratio, varies by patient)
• ~1 Hz: PPG update rate (displayed SpO2 refresh)

Decomposing the signal: AC and DC components

The photodetector current at each wavelength is split into two components by the device firmware:

• DC component: the slowly-varying or constant baseline — light absorbed by skin, subcutaneous fat, bone, venous blood, melanin, and the non-pulsatile fraction of arterial blood. This is the "noise floor" the useful signal sits on top of.

• AC component: the small ripple synchronized with the cardiac cycle — as arterioles distend with each systolic pulse, path length through arterial blood increases slightly, absorbing a bit more light. This ripple, typically only 1-2% of the DC level in healthy perfusion, is the only part of the signal that reflects true arterial oxygenation.

Bandpass filtering (roughly 0.5–5 Hz, matching physiological heart rates) isolates the AC component from motion artifact and ambient light interference, then the peak-to-trough AC amplitude is normalized by the DC level to make the measurement independent of overall tissue thickness and LED brightness.

How melanin shrinks the usable signal

Melanin sits almost entirely in the non-pulsatile epidermal layer, so its main effect is to inflate the DC baseline without contributing to the AC pulsatile ripple. Two consequences follow directly:

• Lower perfusion index: since perfusion index ≈ AC/DC, a larger DC denominator from melanin absorption mechanically shrinks the calculated perfusion index — even when true arterial pulsation is completely normal. Devices sometimes flag "low signal quality" more often in highly pigmented skin, though many consumer devices display a reading anyway rather than rejecting the measurement.

• Higher gain, higher noise: to keep the detector within its dynamic range despite melanin absorbing much of the emitted light, automatic gain control increases photodiode amplification. This proportionally amplifies electronic and shot noise along with the tiny AC signal, degrading the effective signal-to-noise ratio used to compute the ratio-of-ratios in the next stage.

A 2005 study by Bickler et al. found that pulse oximeters produced more scatter (measurement variability) in subjects with darker skin — not just a shifted average, but a wider spread of possible erroneous readings, particularly during controlled desaturation to SpO2 levels below 90%.

From R-Value to Displayed SpO2 — an Empirical Calibration Curve

Pulse oximeters do not calculate SpO2 from first-principles physics alone. They compute a normalized ratio called R, then look up the corresponding SpO2 value on a calibration curve empirically derived by exposing human volunteers to controlled, brief hypoxia while measuring true arterial saturation with a blood gas co-oximeter. The composition of those original calibration cohorts matters enormously.

• R = (ACred/DCred): Ratio formula (÷ (ACir/DCir))
• ~0.4: Typical R at 100% SpO2 (device/algorithm dependent)
• ~1.0: Typical R at 85% SpO2 (red/IR absorption roughly equal)
• 1970s–90s: Calibration cohort era (predominantly light-skinned subjects)

Computing R and mapping it to a saturation value

The core algorithm, unchanged in essence since the 1970s:

1. Extract AC and DC amplitude for red (660nm) and infrared (940nm) channels over a rolling window of several cardiac cycles 2. Compute R = (AC_red / DC_red) / (AC_ir / DC_ir) 3. Map R to %SpO2 using an empirical calibration curve — approximately linear in practice: SpO2 ≈ 110 − 25×R for many implementations, though exact coefficients are proprietary per manufacturer 4. Smooth and average across multiple pulse cycles, then display a value updated roughly once per second

This calibration curve cannot be derived theoretically with sufficient accuracy — it is built empirically by desaturating healthy volunteers to specific SpO2 levels (as low as 70-80%) under controlled hypoxic breathing conditions, drawing arterial blood samples for gold-standard co-oximetry, and fitting R against true SaO2.

The calibration cohort problem

Because deliberately desaturating human volunteers to hypoxic levels carries real risk, calibration studies have historically used small cohorts — often fewer than 10-30 healthy young volunteers per device model. Regulatory clearance requirements through the 1990s and 2000s did not mandate diverse skin pigmentation representation, and many landmark calibration studies (including some still referenced in current device validation) enrolled cohorts that were predominantly light-skinned.

If the R-to-SpO2 mapping is fit primarily on data where melanin-related attenuation is minimal, the resulting calibration curve implicitly assumes an optical environment that does not match darker-skinned users. The algorithm has no way to detect or correct for this mismatch at the point of use — it simply applies the same curve to every user, regardless of skin tone, because the device has no independent measurement of melanin content.

Feiner, Severinghaus, and Bickler (Anesthesia & Analgesia, 2007) directly tested pulse oximeters during controlled hypoxia in light- and dark-skinned subjects and found clinically significant overestimation of true saturation in darker-skinned participants, worsening as true SaO2 fell below 90% — the exact pattern later confirmed at population scale by Sjoding et al. in 2020.

Occult Hypoxemia — When a Normal-Looking Reading Hides Real Danger

The clinical stakes of this optical bias became impossible to ignore during the COVID-19 pandemic, when pulse oximetry was used at massive scale to triage patients for supplemental oxygen and hospital admission. Large retrospective studies revealed that Black patients were substantially more likely than white patients to have "occult hypoxemia" — a pulse oximeter reading of 92% or above while their true arterial oxygen saturation, measured by blood gas, was actually below 88%.

• ~3x: Occult hypoxemia, Black patients (more frequent than white patients)
• >10,000: Sjoding et al. cohort (paired SpO2/SaO2 measurements)
• 2024: FDA revised guidance (draft update on accuracy/labeling)
• ≤3.0%: Required Arms accuracy (RMS error, diverse skin tones)

The landmark evidence: Sjoding et al., NEJM 2020

The pivotal study by Sjoding and colleagues (New England Journal of Medicine, December 2020) analyzed over 10,000 paired pulse oximetry and arterial blood gas measurements from two large hospital systems. Its central finding: among measurements where the pulse oximeter read 92-96% (a range often considered reassuring, not requiring urgent intervention), Black patients had nearly three times the frequency of occult hypoxemia — true SaO2 below 88% — compared to white patients.

A companion 2022 study (Fawzy et al., JAMA Internal Medicine) analyzing COVID-19 patients found that occult hypoxemia measured this way was associated with Black and Hispanic patients receiving supplemental oxygen and COVID therapies later than white patients with equivalent true hypoxemia — because their pulse oximeter readings looked falsely reassuring to treating clinicians, delaying escalation of care.

These are not laboratory curiosities: occult hypoxemia at this scale, embedded in a device used for triage across millions of patient encounters per day, translates into a measurable disparity in the timeliness of oxygen therapy, ICU admission, and other time-sensitive interventions.

Regulatory response — FDA guidance and testing requirements

Following an FDA advisory panel meeting in November 2022 that publicly examined this evidence, the agency began revising its expectations for pulse oximeter clearance:

• Draft guidance (2024) proposes that premarket clinical accuracy studies enroll participants across a range of skin pigmentation — not just self-reported race, but objectively measured melanin content (e.g., via a calibrated skin tone scale) — with a minimum number of participants in dark-pigmentation categories • Proposed accuracy standard: Arms (root-mean-square accuracy, comparing device SpO2 to co-oximetry SaO2 across the desaturation range down to ~70%) should not exceed roughly 3.0%, evaluated separately across pigmentation subgroups rather than only in aggregate • Updated labeling requirements aim to disclose known accuracy limitations, including performance variation by skin pigmentation, directly to clinicians and, increasingly, to consumers • The guidance explicitly extends discussion to consumer wearables (smartwatches, rings) marketed with SpO2 features, which have historically faced lighter regulatory scrutiny than FDA-cleared medical fingertip pulse oximeters

These changes remain in progress; as of the 2024 draft, manufacturers are not yet universally required to report pigmentation-stratified accuracy data, and most currently marketed consumer wearables do not publish such breakdowns.

The core engineering fix under discussion is not exotic: expand calibration study cohorts to include a representative range of Fitzpatrick skin types at every desaturation level tested, and report accuracy (bias and RMS error) separately per skin-tone subgroup rather than as a single pooled number that can mask large subgroup-specific errors.

What consumers and clinicians can do today

Until devices and their calibration datasets improve, several practical mitigations are widely recommended:

• Treat pulse oximeter readings as one data point, not a definitive measurement — especially in the 88-96% range, where the clinical action threshold sits closest to where bias is largest • Weight trends over single readings: a consistent downward trend in wearable SpO2 is informative even if the absolute number carries some bias • In clinical settings, corroborate borderline pulse oximetry readings with arterial or venous blood gas measurement before ruling out hypoxemia in patients with darker skin pigmentation • Consumer wearable SpO2 features are generally marketed as "wellness" indicators, not diagnostic-grade measurements — a distinction that carries real weight given the documented bias, but is often lost on users who rely on the number for health decisions

Approximate measured-vs-true SpO2 bias by Fitzpatrick skin type (illustrative, at true SpO2 ≈ 85%)