Towards a Wearable Skin Tone Responsive Optical Sensor ^†

Abstract

Melanin is one of the key light absorbers in skin and is responsible for the colour of the skin. This study evaluates the responsivity of different skin tones to white light within the visible spectral range of 300–700 nm on 12 participants. The results show that the peak amplitude of the reflected light signal decreased by 90% for darker skin tones, compared to 70% for lighter skin tones. There were also visible differences at the 460 nm and 570 nm wavelengths between the skin tones, suggesting that the standard one-glove-fits-all pulse oximeter might not be ideal.

1. Introduction

Skin plays a pivotal role in the development of wearable optical sensors. The aftermath of COVID-19 brought about realization of the inaccuracies of pulse oximeters in measuring the peripheral oxygen saturation levels $\left(Sp{O}_2\right)$ in darker-skinned patients [1]. Sdjoding et al. [2] demonstrated that 88 out of 749 (11.7%) dark-toned patients were misdiagnosed as having normal $Sp{O}_2$ levels ranging between 92% to 96%, while they were actually experiencing occult hypoxemia. The same tests showed that 99 out of 2778 (3.6%) of light-toned patients at the University of Michigan Hospital suffered from a similar diagnostic error.

The main chromophores responsible for the behaviour of light on the skin are water, melanin, and haemoglobin [3,4,5,6]. Melanin, which is responsible for the pigmentation of the skin, or the skin tone, is recognised as the main absorber of light [7]. The interaction of oxygenated and deoxygenated haemoglobin with light is used in the estimation of $Sp{O}_2$.

Green, red, and infrared (IR) light-emitting diodes (LEDs) are commonly found in commercial optical wearable devices. In [8], an in-depth discussion is given on light sources, including their skin penetration depths. They used a combination of red and green LEDs to improve the detection of their PPG sensor. Another study [9] used red and IR LEDs with a concave housing structure and a phototransistor as the detector for finger monitoring. The calculation of $\left(Sp{O}_2\right)$ in these articles was based on the ratio (R) of the red and green LEDs using their AC and DC values, referring to the pulsatile and non-pulsatile light absorption characteristics of the arteries and tissues present, and their molar extinction coefficient ($ϵ$) [8,10]. Kim et al. [11] suggested that although theoretically, skin colour does not affect the accuracy of the $\left(Sp{O}_2\right)$ measurement, the crosstalk which occurs at the surface of the skin influences the AC/DC ratio, leading to errors in the calculation of the $\left(Sp{O}_2\right)$ value. From the Beer–Lambert Law [8,10], $\left(Sp{O}_2\right)$ is affected by the ratio of absorptivity of the light sources, R.

In 2020, Ellis Monk introduced a 10-shade scale as an improvement to the Fitzpatrick Scale, which has been widely used as the gold standard to classify different skin tones over the years [12]. The previous 6-shade scale was suggested to have overlooked some skin tones, leading to occasionally incorrect identifications when used in artificial algorithms for facial recognition. The colour of an object, including skin, can be expressed numerically through colour science. The Commission Internationale de l’Éclairage (CIE) XYZ color space was adopted in this study to classify the volunteers using an objective method as opposed to the popular subjective method of visual assessment and categorisation of skin tone to match the Fitzpatrick scale.

This research observed the reSpOnsivity of 12 different skin tones to a white light source within the visible spectral range of 300–700 nm. It was noted that darker-skin tones responded differently to wavelengths within the indigo-blue band and the yellow-orange band of the spectrum as compared to lighter-skin tones.

2. Materials and Methods

2.1. Participant Selection

Twelve individuals with an even distribution of light to dark tones, comprising six females and six males within an age range of 20 to 50, were randomly selected from the University of Pretoria. Each of the participants was notified of the risks involved with the experiments as guided by the Ethics Committee of the University of Pretoria, and indemnity forms were signed for consent.

2.2. Instruments

A white light-emitting diode (LED) with a frequency range of 300 to 700 nm was used as the light source for all the experiments. The initial series of reflectance measurements was conducted utilising a Starline AvaSpec 2048 Spectrometer, equipped with a CC-VIS/NIR cosine corrector connected to the FC-UVIR600-2-BX optical cable; all components were manufactured by Avantes, Apeldoorn, The Netherlands. This setup was then changed to an 11-Channel Adrafruit AS7341 spectral colour sensor circuit, which was printed on a flexible Kapton substrate. An Arduino Uno R3 board was used as an interface between the sensor and the Python, version 3.12 script for processing the data.

The reflected white light from the skin of the volunteers was measured using the two experimental setups, as seen in Figure 1. The optical sensors were placed on the inner wrist of the participants, which appeared to be less exposed to the tanning effects of the sun. The wrists were cleaned with alcohol wipes on the area of interest to ensure that the surface was free from debris, as well as body lotion or creams, which could influence the behaviour of the illuminated light.

All measurements were carried out in a dark laboratory to minimize ambient light interference. The measurement setups were housed in a 3D-printed case, which was developed using the software Fusion 360 and printed on a stereolithographical Formlabs 3/3A printer using black resin. The casing ensured that for all readings taken, the photodetector was 1 cm from the surface of the skin, as well as 1 cm apart from the light source. The room temperature measured between 19.5 and 20.5 °C.

3. Results

3.1. Classification of Skin Tones

Figure 2 illustrates the 12 skin tones which were measured as well as the CIE XYZ colour output. It categorizes the skin tones into the corresponding Monk scale equivalents.

3.2. Spectrum Analysis of Different Skin Tones

The normalised results of the reflectance measurements using both the commercial spectrometer and the AS7341 sensor manufactured by Avantes, Apeldoorn, The Netherlands showed that lighter skin tones reflected light in the shorter wavelengths of 450 nm and absorbed more light in the longer wavelengths between 575 and 650 nm. The opposite was noted in the darker skin tones. Although these wavelengths are subject to haemoglobin, the resulting pattern in Figure 3 identifies a distinct grouping according to skin tone, indicating that there is a presence of influencing factors at those wavelengths other than just haemoglobin.

4. Discussion

The CIE XYZ values measured provided a distinction between the skin tones of the participants, enabling them to be grouped according to light tones (CIE XYZ = 0.4, 0.3, 0.08) and darker tones (CIE XYZ = 0.2, 0.1, 0.03). The accuracy of the colour was influenced by the red shade which appeared on the surface of the skin when the white LED was illuminated 1 cm from the surface. This was more evident in lighter skin tones. Further data manipulation is required to remove this effect to obtain melanin information aligned to the skin colour scale.

Although the Monk Scale is a welcome upgrade to the Fitzpatrick scale, it still faces limitations as it provides discrete tones which do not apply in real-life situations. An electronic definition of skin tone through wavelength fingerprinting could provide a broader range of skin colour, increasing the accuracy of recognition patterns.

The study showed that darker skin tones reflected more light between wavelengths of 550 and 650 nm, producing a flatter gradient as compared to lighter tones, which reflected more at the blue wavelength of 450 nm. The shorter wavelength absorption of darker tones could be explained by the presence of an increased concentration of melanin in the skin, resulting in more photons being absorbed. However, a larger sample size, inclusive of a wider variety of skin tones, would provide a more detailed analysis.

Since $\left(Sp{O}_2\right)$ is calculated using the ratio, R, of absorptivity of red light over the absorptivity of green light, if the red light DC, factoring in the effect of skin pigmentation, decreases, R decreases, resulting in an overall increase in the final $\left(Sp{O}_2\right)$. Apart from the crosstalk mentioned in the introduction, this could be the cause for the exceptionally higher $\left(Sp{O}_2\right)$ values in darker-toned individuals. The change in DC is influenced by skin pigmentation.

The manufacturing readiness level of this research is at level 3, whereby laboratory experiments were conducted to provide proof of concept. Future work will involve using these findings in developing a personalised sensor which would improve the accuracy of pulse oximeter diagnosis.