Light source spectra and higher measurement uncertainty in pulse oximeter readings for individuals with darker skin pigmentation

Editor—Skin colour bias in pulse oximetry has been reported^1,2 such that pulse oximeter measurements (peripheral blood oxygen saturation [SpO₂]) overestimate the true arterial oxygen saturation (SaO₂) of individuals with dark skin, leading to undiagnosed hypoxaemia. We hypothesised³ that this bias resulted from the optical properties of the two light-emitting diode (LED) light sources, one red and one infrared (IR), typically used in commercial pulse oximeters. The detector measures the amounts of red light and IR light transmitted through the finger over each pulse cycle. Changes in red and IR light transmissions through arterial blood are isolated by normalising them to their mean transmissions over the pulse cycle. The ratio of red over IR changes in normalised transmission (R) varies inversely with blood oxygenation.

To make pulse oximetry measurements meaningful to clinicians, R values are converted to SpO₂ values based on an empirical calibration relative to simultaneous SaO₂ measurements. Skin colour bias arises from the polychromatic LED sources. These sources are not monochromatic, having full-width-at-half-maximum emissions of 20–60 nm. Our hypothesis, based on the Beer–Lambert law,⁴ was that the skin-colour bias is caused by a melanin-dependent wavelength shift of the transmitted polychromatic light spectral distributions; the greater the melanin concentration in the skin, the greater the wavelength shift. Our study confirmed this hypothesis by comparing R values derived from narrowband (nearly monochromatic) and polychromatic sources, showing that skin colour bias could be eliminated by using monochromatic red and IR light sources.⁵

Some studies have suggested that, in addition to the overall melanin bias, there is a systematic relationship between skin colour and SpO₂ measurement uncertainty, with uncertainty being greater for subjects with darker skin pigmentation.^1,6 Although the Beer–Lambert law can explain the source of melanin bias, it does not accurately model melanin distribution within the skin. Inferences based upon a simple application of the Beer–Lambert law assume that spectrally absorbing substances, such as melanin, are uniformly distributed throughout the medium, as if suspended in solution. In fact, discrete, spectrally absorbing melanin molecules (epidermal melanin units [EMUs]) are not evenly distributed throughout the epidermal layer of the skin.⁷ Moreover, the epidermal layer is not strictly planar but exhibits convoluted structures, as can be seen in images of stained skin sections.^8,9 The discrete, nonuniformly distributed EMUs and textured boundaries of the epidermal layers strongly suggest that repeated SpO₂ measurements from the same nominal area of the skin (e.g. fingertip) would not always be identical because the light source emission–detection path lengths would vary. If this were true, attenuation of the different optical path lengths through a different number of EMUs in the epidermal layer would contribute to greater variability in SpO₂ measurements.

We therefore hypothesised that the observed melanin bias in SpO₂ measurements would necessarily be associated with measurement uncertainty. If this hypothesis were true, it would logically follow that the variability in independent measurements of SpO₂ (based on R) would increase with overall melanin concentration in the skin, and also that measurement variability would be unrelated to R for monochromatic light sources.

We tested this hypothesis in eight of the 31 subjects from our previous study (four males and four females)⁵ selected for repeated, independent R measurements. These subjects represented a range of skin pigmentation from ‘light’ to ‘brown’ by individual topography angle (ITA) categorisation. Four subjects self-identified as White, three as Black, and one as Asian. The ages of the subjects ranged from 22 to 56 yr (mean age 38.1 yr). Two subjects were authors (KB and AB). As before, subjects underwent a controlled oxygen desaturation procedure while data were continuously collected using spectrally resolved photoplethysmography (srPPG).⁵ The procedure was performed on different days to permit determination of measurement uncertainty from multiple measurements and to ensure that these measurements were independent. Seven subjects completed three sessions, and one completed two sessions.

We used the techniques and metrics developed for our previous study.⁵ The srPPG technique uses a custom fingertip oximeter with a broadband light source and a spectrometer to collect pulsatile waveforms for a wide, continuous range of transmitted wavelengths. To compare the effects of spectral bandwidth on the calculation of R, we applied digital spectral bandpass filters to the srPPG spectral transmission data. This allowed us to calculate R for a narrowband light source and R for a polychromatic light source using the same spectral transmission data. The functional relationship between the two sets of R values (R-slope) for each experimental session was determined using least squares linear regression. This R-slope metric represents the bandwidth sensitivity of the subject's skin; an R-slope of unity indicates no bandwidth-related sensitivity, whereas a negative R-slope indicates bandwidth-related sensitivity. R-slope was previously found to correlate with melanin content; the more melanin in the skin, the more negative the R-slope value.⁵

To quantify the measurement-to-measurement uncertainty, the standard deviation of the R-slope measurements for every session (Š_R-slope) was calculated for each subject. Melanin fraction⁵ was used to quantify the concentration of melanin in a subject's finger. White light is transmitted through the fingertip, and the spectral transmission data are collected. These data are then modelled using published spectral transmission curves of the various constituents of the fingertip (water, bone, fat, melanin). Melanin fraction represents the amount of melanin needed in the model to predict the overall spectral transmission data relative to the total of all constituents. Melanin fraction correlates well with ITA, another method of estimating melanin based on skin spectral reflectance.¹⁰

The Š_R-slope was normalised by dividing by the mean R-slope value for each subject to calculate the coefficient of variation for each subject. The R-slope coefficient of variation was found to correlate significantly with the mean melanin fraction (Fig. 1, Pearson's correlation coefficient [r]=0.843, P=0.0085). Subjects with a higher melanin fraction had greater R-slope coefficient of variation values. Although melanin fraction (based on transmission) was shown⁵ to be a better metric than ITA (based on reflectance) for the characterisation of skin pigmentation in this application, we nevertheless compared the coefficients of variation across light- and dark-skinned subjects, with subjects divided into tertiles based on ITA. We compared the mean coefficient of variation of the three subjects with the highest ITA values (lightest skin) to the mean coefficient of variation of the three subjects with the lowest ITA values (darkest skin). As predicted based on the melanin fraction analysis, the mean for low ITA (0.0324) was higher than the mean for high ITA (0.0041) (P=0.027 by one-tail t-test).

Fig 1.

Coefficient of variation of R-slope values vs melanin fraction for eight subjects. Melanin fraction is the modelled attenuation of the spectral transmission of white light through the subject's finger caused by the melanin concentration. R-slope values are determined by plotting the R values for broadband sources as a function of the R values for narrowband sources obtained simultaneously. The standard deviation of R-slope values (Š_R-slope) for each subject was determined from repeated independent measurements. The coefficient of variation was calculated as Š_R-slope divided by the mean R-slope for each subject. The best fitting, least-squares, linear relationship between the coefficient of variation and melanin fraction means (solid discs, error bars = 1 standard deviation), illustrated by the dashed line, and the coefficient of determination (R²) are shown.

Therefore, melanin concentration appears to have the same effect on SpO₂ measurement uncertainty when using polychromatic emitters as it does on measurement bias. Although this pilot study does not conclusively prove this, our findings show the importance of further investigation into the effect of oximeter light source bandwidth on melanin measurement uncertainty and melanin measurement bias. These preliminary findings, if confirmed using narrowband light sources, indicate that dark-skinned individuals are doubly at risk for misdiagnosis of clinically significant hypoxaemia because of both overall measurement bias and measurement uncertainty.

Authors' contributions

Study conception, principal investigator and guarantor: MSR

First draft preparation: KB

Data analyses: AB, KB

Drafting, revision and approval of the manuscript: all authors

Approval of the final manuscript: all authors

Funding

US National Institute of Biomedical Imaging and Bioengineering (R01EB033799) to extend their theoretical and empirical analyses and to develop a pulse oximeter using narrow spectral bandwidth light sources.

Declaration of interest

The authors declare that they have no conflict of interest.