Spectral emission profile and wavelength tolerances affect pulse oximeter performance
University of Freiburg, Faculty of Engineering, 79110 Freiburg, Germany
HAW Landshut, Faculty of Electrical Engineering, 84036 Landshut, Germany
Hahn-Schickard, 79110 Freiburg, Germany
✉Correspondence: maximilian.reiser@haw-landshut.deAbstract
We investigate the effects of skin pigmentation and light source characteristics on the performance of reflective Pulse oximetry (PO) devices used in healthcare and well-being applications. We use Monte Carlo (MC) simulations to compare ideal monochromatic and realistic LED spectral emission profiles and tolerance-related wavelength shifts. The simulation covers photon transport in skin models with melanin concentrations (2.55% to 30.5%) and arterial oxygen saturations SaO2 (70% to 100%.) Accuracy was assessed by SpO2 error, root-mean-square error RMSE (Arms), and percentile tail-errors (P90, P95, and P99).
Monochromatic spectral emission yielded the lowest SpO2 error (RMSE = 1.32), while LED spectral emission profiles increased errors (RMSE = 2.10). Infrared wavelength tolerances increased SpO2 RMSE by 1.1 ± 0.3. SpO2 error increased with melanin concentration, from underestimation (−1.8 ± 0.1%) at 2.55% melanin concentration to overestimation (+3.9 ± 1.2%) at 30.5% for low SaO2 (70%) and LED spectral emission profiles. At 30.5% melanin concentration, P95 and P99 exceeded FDA and DIN EN ISO 80601-2-61 thresholds, in particular at low SaO2 (70%). Clipping SpO2 estimates at 100% resulted in an apparent RMSE decrease of up to 3%, reflecting error masking rather than real error reduction.
In conclusion, LED spectral emission profiles and wavelength tolerances can amplify melanin-related bias in SpO2 estimates. Monochromatic emission and tighter wavelength control can reduce SpO2 error and should be considered in device design and regulation. Regulatory standards should discourage clipping SpO2 estimates at 100% and mandate additional metrics as RMSE fails to reflect clinically critical percentile error thresholds, i.e. P95 and P99.
Article notes
Competing Interest Statement
The authors have declared no competing interest.
Introduction
Photoplethysmography (PPG) is a widely used non-invasive optical method to measure pulsatile changes in blood volume caused by the heartbeat. Reflective PPG measurements are increasingly relevant due to their integration into wearable devices and their ability to continuously monitor physiological parameters (1–3). Pulse Oximetry (PO) is a clinically important application of PPG, which estimates arterial oxygen saturation SaO2 by comparing absorption at two wavelengths in the red and infrared spectrum. For PO, light is emitted into tissue and reflections are detected by a photodetector.
The PPG signal consists of two components, the pulsatile (AC) component and the static (DC) component. Static absorbers include skin, bone, and the constant portion of blood volume, while the pulsatile component reflects changes in arterial blood flow. The perfusion index (PI), defined as the ratio of AC and DC components, provides a measure of peripheral perfusion and vascular tone. The ratio of ratios (RoR), defined as the ratio of the PI at the red wavelength to the PI at the infrared wavelength, is calibrated against reference measurements of arterial oxygen saturation SaO2.
Skin pigmentation is a critical factor for SpO2 performance, as PO tend to overestimate oxygen saturation in individuals with darker skin tones, i.e. increased melanin concentration (4, 5). Skin with increased melanin concentration (i.e., 30.5%) shows a larger absorption coefficient µa, compared to lighter skin (i.e., 2.55%), with non-linear effects across red and infrared wavelengths. The non-linear signal distortion due to melanin concentration results in a systematic overestimation of oxygen saturation (5), in particular at low levels of arterial oxygen saturation SaO2. SpO2 estimation error can have serious clinical consequences, in particular when arterial oxygen saturation SaO2 is below critical values (e.g., < 88% (6, 7)), but the PO device erroneously indicates clinically acceptable levels of SpO2 (called occult hypoxemia) (8). The discrepancy in SpO2 estimation may delay oxygen therapy or other urgent interventions, especially in darker-skinned patients, who are disproportionately affected. For instance, studies have shown that patients with increased melanin concentration (e.g., 30.5%) are up to three times more likely to experience undetected hypoxemia than patients with low melanin concentration (2.55%) (9).
The accuracy of PO devices is assessed using the root-mean-square error (RMSE, Arms) between a reference oxygen saturation value SaO2 and the SpO2 estimation by the PO device. Reference SaO2 is obtained through invasive blood gas analysis (10, 11), which is considered as gold standard. However, certified PO devices may still exhibit systematic errors (9, 12, 13), which remain undetected when relying solely on RMSE. Since RMSE only reflects the global average deviation, it does not necessarily capture systematic biases in specific subpopulations, e.g., individuals with increased melanin concentration or low arterial oxygen saturation. In addition, manufacturers often clip SpO2 estimations at 100% oxygen saturation, which alters and reduces the subsequently calculated RMSE by masking potential overestimation errors.
For PPG and PO measurements, typically light emitting diodes (LEDs) are used as a light source. LEDs are specified by a nominal wavelength (e.g., 660 nm and 940 nm for PO), usually denoting the centroid of their wavelength emission profile. In reality, LEDs are not monochromatic but emit over a spectral band, e.g., with typical full width at half maximum (FWHM) between 17 nm and 42 nm (SFH 7014C by ams-OSRAM AG). The wavelength spectral emission profile arises from semiconductor band-to-band recombination and thermal carrier distributions. In addition, LEDs are subject to wavelength tolerances, which originate from bandgap variations (including composition, doping, and crystal structure) and manufacturing processes (including epitaxy, layer thickness, defects, and wafer-to-wafer variation). In addition to the intrinsic spectral emission profile, manufacturing variations introduce nominal tolerances ranging from ± 2.5 nm to ± 11.5 nm (e.g., ams-OSRAM SFH 7014C) relative to the specified nominal wavelength. The selection of wavelength is critical in PO because the absorption of oxyhemoglobin O2Hb and deoxyhemoglobin HHb change steeply in the red wavelength spectrum (14). Even small shifts can therefore alter the RoR and error in SpO2 estimation. Milner and Mathews (15) reported deviations up to 7% for ± 2 nm shifts in the red spectrum, while Rea et al. (16) and Bierman et al. (17) showed that LED spectra contribute to skin-pigmentation bias, which can be mitigated by narrower-band emitters.
Monte Carlo (MC) simulations are a tool to isolate and quantitatively capture complex spectral effects of light sources in PPG and PO (5, 18–20). MC simulations can provide a physically accurate framework to model photon–tissue interactions, including absorption, scattering, and anisotropy, which cannot be captured by simplified Beer–Lambert or diffusion-based approaches. MC simulations can integrate spectral emission profile, detector responsivity, layered and heterogeneous tissue structures (including epidermis, dermis, subcutaneous tissue, bone, or blood vessels), and physiologically relevant changes in blood volume. MC simulations are the gold standard for simulating photon transport in biological media with scattering.
Beyond reproducing the fundamental AC and DC components of the PPG signal, MC simulations offer the chance to analyse photon path length distributions, penetration depth, or PI. Spectral sweeps can be handled efficiently by computing tissue responses across a wavelength grid and applying source and detector spectral weighting in post-processing. Recent studies have successfully employed MC simulations to investigate dual-wavelength PPG formation and the impact of melanin on SpO2 estimation error (5, 18, 21).
Milner and Mathews (15) highlighted that nominal wave-length shifts can induce large SpO2 estimation errors, yet did not account for spectral emission, scattering, or melanin. Tsiakaka et al. (22) used a simplified absorption model to optimise wavelength pairs, assuming monochromatic sources and homogeneous tissue. Bierman et al. (17) demonstrated experimentally that LED spectral emission increases melanin-dependent SpO2 estimation error, but neglected manufacturing tolerances, beam profiles, and additional metrics, including Arms. In contrast, our study systematically disentangles the independent effects of spectral emission profiles and wavelength tolerances, which to date have either been considered separately or with substantial simplifications.
In this work, we investigate wavelength tolerances and spectral emission profiles of commercially available LEDs to reflect realistic device conditions. Melanin concentrations were selected to represent Fitzpatrick scale skin types (23), and SaO2 values were aligned with international PO standards (10, 11). In addition, source-detector distance was varied to disentangle geometry influences from PO performance. Empirical SpO2 estimations are sensitive to physiological variability and other non-controllable parameters, including peripheral perfusion of individuals and temperature-dependent wavelength shifts. By employing MC simulations, we incorporate realistic tissue heterogeneity (epidermis, dermis, subcutaneous fat, muscle), melanin concentrations, detector responsivity, and photon scattering, modelling a physical representation of the two physiological phases (i.e. diastole and systole) in PPG formation. Unlike previous work, we quantify the spectral effects not only via absolute SpO2 estimation error (RMSE, Arms), but analyse the RMS deviation of SpO2 estimation as well as tail-errors. Furthermore, by analysing wavelength tolerance and spectral emission profile under identical simulation conditions, we provide the first direct comparison of their relative impact. Our approach allows us to derive actionable design insights for light source specifications, source–detector geometry, and PO device calibration strategies that remained unaddressed so far.
In summary, this paper provides the following contributions:
- We perform realistic MC simulations using a validated framework to analyse SpO2 estimation performance in wearable and mobile health devices. Our analysis covers tissue heterogeneity, melanin concentrations, physiological phases, detector responsivity (spectral and angular), light source spectral and angular emission profiles, nominal wavelength tolerance, and source-detector distance variability, to reflect conditions encountered in wearable SpO2 monitoring.
- We introduce an extended error analysis beyond absolute SpO2 estimation error that includes tail-error quantification.
- We derive actionable design insights for light source specifications, source-detector geometries, and calibration strategies for pulse oximeters.
Methods and Materials
We deployed a framework consisting of three main components: (1) combined skin and sensor model, (2) photon-skin simulation, and (3) reflective PO simulation (see Fig. 1). The skin model included layer-specific optical characteristics (absorption coefficient µa, scattering coefficient µs, anisotropy factor g, and refractive index n), anatomical characteristics (skin layers, thickness), melanin concentration, arterial oxygen saturation SaO2, and physiological states (systole and diastole). The sensor model included the spectral and angular sensitivity of the photodiode and angular and spectral emission profile of both the LED and an assumed ideal monochromatic light source. The photon-skin simulation propagated photon packets, based on the sensor model, through our skin model. For each configuration, we recorded the detected intensity I, photon packet count, absorption per skin layer, and computation time across source-detector distances ranging from 3.5 mm to 11.5 mm.
A.Sensor model
As a light source, we modelled the LED SFH7014C and as a detector the photodiode SFH2704A (both by ams-OSRAM AG). The spatial angular emission profile was represented by the relative source intensity Irel (see Fig. 2).
For tolerance analysis, we varied the nominal wavelengths within their specified tolerance ranges, i.e., 655 nm ± 2.5 nm for the red spectrum and 940 nm ± 9.5 nm for the infrared spectrum. For LED spectral emission simulations, we simulated spectral emission at FWHM in the red spectrum of 17 nm and in the infrared spectrum of 42 nm. Tab. 1 shows the relative spectral emission Ie,rel defined at discrete support points. For spectral emission simulations of the monochromatic light source, we assumed an ideal spectral emission profile and simulated the nominal wavelength in the red and infrared spectrum (655 nm and 940 nm).
The detected Intensity I was adjusted to the relative spectral sensitivity of the photodiode SS,λ and the directional characteristics Srel,φ.
B.Skin model and photon–skin simulation
We used a previously developed multilayer skin model which consists of six skin layers (epidermis, capillary loops, upper plexus, reticular dermis, deep plexus, and hypodermis), and an additional muscle layer (5). We simulated both physiological states, systole and diastole. During systole, the blood volume fraction was increased in skin layers (capillary loops, upper plexus, reticular dermis, deep plexus, and hypodermis) (19). The photon-skin simulations were performed using a previously developed MC framework for photon-tissue interactions, implemented in C++ and CUDA (24). For each parameter configuration (defined by physiological state, wave-length λ, monochromatic or LED spectral emission profile, melanin concentration CMel, and arterial oxygen saturation SaO2) a total of 5 × 109 photon packets were simulated. Each configuration required an average of 234.1 ± 22.8 s of processing on an NVIDIA RTX 6000 Ada.
C.Reflective pulse oximetry simulation
Oxygen is transported bound to hemoglobin in erythrocytes. The arterial oxygen saturation SaO2 is defined as the fraction of O2Hb relative to the total functional hemoglobin (O2Hb + HHb). Hemoglobin absorbs light differently depending on its oxygenation state. The absorption coefficient µa of O2Hb is greater in the infrared wavelength spectrum compared to the red wavelength spectrum, with approximately 4.3-fold higher absorption at 940 nm compared to 655 nm. (14). In contrast, HHb increases absorption in the red wavelength spectrum compared to the infrared spectrum, with approximately 3.8-fold greater absorption at 655 nm compared to 940 nm (14). The opposing absorption properties of O2Hb and HHb in the red and infrared spectrum can be used to empirically estimate oxygen saturation SpO2.
For estimating SpO2, PI was derived as the ratio of the pulsatile to the non-pulsatile signal components:
Based on the PI in the red and infrared spectrum, RoR was calculated as the relative relationship between the distinct absorption characteristics of O2Hb and HHb:
Using a manufacturer-specific empirical calibration, the arterial oxygen saturation SaO2 was estimated from RoR as: where A, B, and C are calibration constants.
PO devices can be classified as medical devices and thus their use is regulated by international standards, including DIN EN ISO 80601-2-61(11), and the U.S. Food and Drug Administration (FDA) (10). To ensure reliability, the regulations specify minimum performance requirements. RMSE, denoted as Arms, quantifies the deviation between the reference oxygen saturation SR (e.g., determined by blood gas analysis) and the oxygen saturation SpO2 estimated by the PO. The RMSE is calculated as:
The RMSE limit required by the FDA for reflective PO is below 3.5% and by DIN EN ISO 80601-2-61 below 4.0%. For each parameter configuration (SaO2, systolic and diastolic state, melanin concentration, source-detector distance, wavelength combination, and monochromatic or LED spectral emission profile), we performed simulations with 25 seeds with 5 × 109 photon packets each. The parameter seeds were used to introduce variability in photon interactions. Different seeds resulted in variations in simulation outcomes (including photon path length and scattering angle). We interpreted the parameter seed as natural interpersonal variability, corresponding to a virtual study cohort.
Based on the virtual study cohort, a general calibration curve was calculated. The deviation between the estimated oxygen saturation SpO2 and the actual input simulation parameter for oxygen saturation (denoted as SaO2 or SR) was then evaluated to determine the RMSE. The calibration and RMSE calculation was repeated for all simulated source–detector distances ranging from 3.5 mm to 11.5 mm.
D.Experiments and validation
In this study, we investigated the impact of two hardware-related factors on reflective PO: (1) manufacturing-related tolerances in nominal LED wavelengths and (2) the spectral emission profile of an LED compared to an assumed ideal monochromatic light source. To quantify the impact in both cases, MC simulations were performed for each scenario. For tolerance analysis, we simulated the endpoints of the specified wavelength ranges (± 2.5 nm at 655 nm and ± 9.5 nm at 940 nm). For spectral emission, both LED and ideal monochromatic light sources were modelled. The approach allowed us to disentangle the independent and combined contributions of tolerances and spectral emission profiles under otherwise identical conditions. The MC framework was validated against laboratory measurements using a porcine skin phantom (24, 25) with angular resolution on both beam incidence and detector detection angles. Furthermore, we validated the MC framework against real-world PPG measurements of systolic and diastolic signal levels from a participant study (26).
The resulting RoR values were calibrated against the reference oxygen saturation (arterial oxygen saturation SaO2 of the skin model), and the estimation accuracy was quantified using the RMSE. The statistical significance of SpO2 estimation analyses was assessed using the paired Wilcoxon signed-rank test (α = 0.05 and α = 0.01), comparing (1) tolerance-related wavelength shifts relative to the 655/940 nm reference and (2) monochromatic versus LED spectral emission profiles. In addition to regulation imposed by the international standards, we analysed tail error quantiles (P90, P95, and P99) of absolute SpO2 estimation error to capture clinically critical deviations in the tails of the error distribution. Increased melanin concentration can systematically influence reflected signal intensities in the red and infrared spectrum, leading to a broader and potentially skewed error distribution in SpO2 estimation. In consequence, individuals with higher melanin levels are more likely to be represented in the extreme error tails rather than by a single deterministic error value.
The analysis was restricted to source-detector distances, representing near-optimal sensor configurations (3.5 mm and 4.5 mm) as a wider range of distances would introduce geometric effects, i.e. alter optimal photon path, which obscure the effects of the investigated variables (i.e., spectral emission profiles, wavelength tolerance).
Results
E.Wavelength tolerance
Fig. 3 shows the SpO2 estimation RMSE for all wavelength tolerance combinations of the red spectrum (652.5 nm, 655 nm, 675.5 nm) and the infrared spectrum (930.5 nm, 940 nm, 949.5 nm). PO calibration was performed based on nominal wavelengths 655 nm and 940 nm, respectively. RMSE increased with source-detector distance. In the infrared spectrum, wavelengths shifts towards 949.5 nm showed worst RMSE compared to shifts towards 930.5 nm. Lowest RMSE was achieved for source-detector distances of 3.5 mm and 4.5 mm.
Fig. 4 shows the SpO2 estimation error (i.e., SaO2 − SpO2) depending on melanin concentration. Here, all wavelength tolerance combinations for both calibrations (655 nm and 940 nm) as well as all SpO2 levels, seeds, and source-detector distances were are aggregated. Lower skin pigmentation (2.55% melanin concentration), yielded an average error of −2.15 ± 1.64%, moderate skin pigmentation (15.5% melanin concentration) an average of −0.56 ± 1.73%, and higher skin pigmentation (30.5% melanin concentration) an average of 1.21 ± 3.31%. With increasing melanin concentration, the average SpO2 estimation error shifts from underestimation to overestimation. Distribution shift in Fig. 4A was caused by the varying source–detector distances. The distribution shift is most pronounced at 2.55% melanin due to higher SNR at low melanin concentrations (27), which helps to resolve distance-dependent differences. With increasing melanin concentrations, increased absorption reduces SNR that blurs the effect of varying source–detector distances.
Wavelength combinations of 657.5/940 nm yielded the lowest RMSE of 1.26, whereas 652.5/949.5 nm resulted in the highest RMSE of 2.64 (see Fig. 5). When infrared wave-length shifted to 949.5 nm, the highest average increase in RMSE of 1.04 ± 0.20 was observed. All deviations relative to 655 nm/940 nm were statistically significant, except for 657.5 nm/940 nm (see Tab. 2).
F.Spectral emission profile
Fig. 6 compares simulated oxygen saturation SpO2 estimates and reference arterial oxygen saturation SaO2. Both, LED and monochromatic spectral emission profiles showed an increasing SpO2 overestimation for high melanin concentrations (30.5%) and oxygen saturation level SaO2 ≤ 80% (see Fig. 6A and B). The monochromatic profile provided more accurate SpO2 estimates (RMSE: 1.32) compared to simulations with the LED profile (RMSE: 2.10). Clipping SpO2 estimations at 100% resulted in systematic overestimation and an apparent reduction of RMSE by 2.9% for the LED profile and by 3.0% for the monochromatic profile (see Fig. 6C and D). All comparisons between the LED and monochromatic emission profiles were statistically significant (see Tab. 3).
Fig. 7 shows absolute SpO2 estimation error distributions quantiles (P90, P95, P99) per melanin concentration and arterial oxygen saturation SaO2. For 2.55% and 15.5% melanin concentration, P90 and P95 remained below the FDA and DIN EN ISO 80601-2-61 thresholds at all oxygen saturation levels. At 30.5% melanin concentration, P95 and P99 frequently exceeded the regulatory limits, in particular at oxygen saturations SaO2 ≤ 80%.
Discussion
Our MC simulation approach allowed us to analyse light source emission profiles and wavelength tolerances under otherwise identical conditions, i.e., without confounding influences (e.g., sensor misalignment, day-to-day physiological variability, and measurement noise). The analysis is vital to guide PO device design considering different skin tones.
Source-detector distances of 3.5 mm and 4.5 mm showed the lowest RMSE. With increasing source-detector distance, photons penetrate deeper tissue layers with larger blood volume fractions and spectral overlap effects (e.g., difference in absorption coefficients µa of O2Hb and HHb) become more pronounced.
Simulations of wavelength tolerances showed that deviations from the nominal wavelength increased the SpO2 estimation error, with RMSE increasing up to 100% relative to the nominal wavelength configuration (see Fig. 5). In our simulations, the wavelength shift from 940 nm to 949.5 nm resulted in the largest RMSE increase from 1.32 to 2.64, primarily due to the increased wavelength-dependent differences in the absorption coefficient of O2Hb and HHb (14). In addition, the increased spectral sensitivity of the photodiode in the infrared range further amplifies the infrared contribution compared to the red wavelength in wavelength tolerance shifts.
Wavelength shifts in the red and infrared ranges are critical, as they can alter the resulting RoR to the extent that RMSE exceeds regulatory limits (FDA and DIN EN ISO 80601-2-61) (see Fig. 3). Therefore, light source and photodiode should be selected with minimal wavelength tolerances. For the light source, it is essential to ensure that wavelength-dependent absorption differences between O2Hb and HHb remain as small as possible in the relevant wavelength regions. For the photodiode, the tolerance-related difference in wavelength-dependent spectral sensitivity should be minimised. Future studies may explore wavelength combinations in the red and infrared spectral range that are minimally influenced by spectral emission characteristics or wavelength tolerances.
Simulated LED spectral emission (FWHM of 17 nm in the red and 42 nm in the infrared spectrum) resulted in an average increase in RMSE of 59.1% compared to monochromatic spectral emission (see Fig.6). The effects of the LED spectral emission profile amplified the dependence of SpO2 estimation performance on melanin concentration, resulting in increased overestimation of SpO2 at higher skin pigmentation (30.5% melanin concentration). In particular, SpO2 overestimation is clinically relevant and can lead to delayed diagnosis and treatment of individuals.
The influence of tolerance-related wavelength shifts and spectral emission profiles on SpO2 estimation are not independent, resulting in a cumulative and asymmetric error distribution (see Fig. 5, 6, and 7). Wavelength tolerances and emission profiles should be restricted to spectral regions, where absorption differences between O2Hb and HHb are minimised. In particular, in the red region around 700 nm and in the infrared region around 920 nm, even minor deviations could substantially amplify SpO2 estimation error.
Our simulations demonstrate that even small wavelength shifts of 2.5 nm can lead to clinically relevant SpO2 estimation bias. Prior work shows that wavelength deviations of 4 nm can lead to an SpO2 estimation error of 2% to 7%, increasing with decreasing oxygen saturation (15). Our results are consistent with literature and further show that wavelength-dependent errors persist under realistic conditions, including spectral emission profiles, varying melanin concentrations, and varying source-detector geometries. We systematically analysed and directly compared the spectral emission profile of light sources and wavelength tolerance shifts under identical conditions. For the first time, we assessed wavelength tolerance shifts and LED spectral emission using RMSE, tail-error thresholds (P90, P95, P99), and connected the results to regulatory requirements (FDA and DIN EN ISO 80601-2-61). Our simulations underline that hardware-level design choices directly influence health equity: The same device may perform within acceptable limits in individuals with light skin tone (e.g., 2.55% melanin concentration), but fail disproportionately in individuals with higher melanin concentrations (e.g., 30.5%). Our results confirm that in order to optimise SpO2 estimation, melanin range calibration might benefit individuals with increasing melanin concentration (5, 28).
SpO2 estimation is influenced by numerous parameters, including beam incidence angle, beam profile, and calibration. In addition, the broad spectral emission profile of LEDs and typical wavelength tolerances can contribute to an overestimation of SpO2 estimation in individuals with increased skin pigmentation (e.g., 30.5% melanin concentration), thereby increasing the risk of occult hypoxemia. The systematic error in SpO2 estimation can contribute to delays in oxygen therapy or inaccurate clinical assessments. Our findings indicate that hardware specifications can help to reduce inequities: (1) narrow emission profile or monochromatic light sources, and (2) more rigorous selection and binning of light sources with respect to wavelength tolerances reduce SpO2 estimation error.
Beyond current standards, which primarily define aggregated error metrics, including RMSE, our findings argue for more explicit spectral specifications in regulations. In particular, maximum allowable FWHM and wavelength tolerances should be required in approval processes, especially for devices intended for a population with varying skin tone. From an ethical perspective, overlooking hardware-level induced errors risk reinforcing disparities in healthcare delivery, in particular in ubiquitous systems, where adoption spans large and heterogeneous user groups.
Furthermore, we highlighted the critical limitation of relying solely on RMSE when evaluating PO performance. While RMSE may suggest compliance to literature, the extreme error tails (see Fig. 7) reveal systematic underperformance with increasing melanin concentration. Exceeding of FDA and DIN EN ISO 80601-2-61 thresholds in P95 and P99 implies that up to 5% of SpO2 estimations could be clinically unreliable in individuals with increased skin pigmentation, in particular at low oxygenation. We conclude that (1) devices that meet the global RMSE criteria may still fail disproportionately in populations with increased melanin concentrations, and (2) reporting percentile error thresholds (e.g., P95 and P99) stratified by skin pigmentation may be necessary to ensure safety in diverse populations. In addition, regulatory standards should prohibit the clipping of SpO2 estimations at 100% as it masks systematic overestimation and artificially improves the apparent performance of PO devices.
While our MC simulations provide a controlled and reproducible setting, they inevitably abstract real-world conditions. The skin model was static and layered, excluding interindividual variability in skin layer thickness, melanin distribution, or vascular density. Motion artefacts, temperature-dependent wavelength drifts, and sensor placement variability were not included, although they are critical in everyday use of wearable devices. However, precisely the abstraction of our simulations highlight the isolated contribution of spectral emission profile and wavelength tolerances, which are difficult to disentangle in-vivo.
Conclusions
We demonstrated that both wavelength tolerances and spectral emission profiles significantly affect SpO2 estimation accuracy (RMSE). SpO2 estimation error increased with increasing melanin concentration, leading to systematic overestimation when arterial oxygen saturation SaO2 declined and thus increasing the risk of occult hypoxemia.
SpO2 estimation errors due to wavelength tolerance shifts and LED spectral emission profiles were additive and in combination could exceed FDA and DIN EN ISO 80601-2-61 regulatory thresholds. For PO device design, narrow-band LEDs or alternative monochromatic light sources combined with minimal wavelength tolerances could reduce SpO2 estimation error and mitigate disparities.
Current regulatory standards rely on RMSE and therefore miss critical tail errors. Percentile error thresholds (P95 and P99) should be considered to properly assess device performance across diverse populations. Furthermore, clipping of SpO2 estimates at 100% may mask systematic overestimation and should be discouraged. Detailed specifications of the sensors used in PO devices (i.e., FWHM, wavelength tolerances) are encouraged.
Overall, PO device accuracy is a multi-parameter optimisation problem (including wavelength tolerances, spectral emission profile, PO calibration strategies, and sensor geometry). Addressing all interacting factors is essential for a consistent PO device performance.
ACKNOWLEDGEMENTS
This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. This work was supported by the Federal Ministry for Economic Affairs and Climate Action (BMWK) on the basis of a decision by the German Bundestag.