A skin-mounted static flux chamber to decompose pulse-oximetry skin-tone bias

Pulse oximeters systematically overestimate blood oxygen in people with dark skin, and this error has real clinical costs: a causal analysis of two intensive-care databases found that oximeter discrepancies kept Black patients on invasive mechanical ventilation 2.3 hours longer in eICU and 7.4 hours longer in MIMIC-IV Path-specific effects for pulse-oximetry guided decisions in critical care. The field explains this bias almost entirely as an optics problem — melanin in the epidermis absorbs the red and infrared light before it reaches the pulsatile vessels — and treats the skin under the sensor as an inert filter. But the sensor site is not inert: it is a living surface with active microvascular gas exchange and a dense, metabolizing microbial community. No one has ever measured either of those directly at the site where the oximeter sits, because the tools to do so have never been brought from environmental science to skin.

This proposal adapts the static chamber — the standard, calibrated method for measuring gas exchange between a surface and the air, long used on soils and composts — to human skin, and pairs it with 16S rRNA gene sequencing (a DNA-barcoding method that profiles microbial communities without needing to culture them) and objective skin-tone measurement. The goal is a first direct answer to a question the literature currently cannot ask: how much of the oximetry bias is optical attenuation by melanin, how much reflects local transcutaneous gas exchange, and how much comes from microbial metabolism on and in the skin itself.

Core Claim

The claim is decompositional and testable. Skin-tone metrics already predict oximetry bias — in a prospective ICU cohort of 128 patients and 521 paired readings, the Monk Skin Tone scale predicted bias at $p=0.01$ and a six-variable objective skin model at $p=0.02$, with a -2.40 percentage-point bias difference between the lightest and darkest Monk tones (95% CI -4.32 to -0.48) Utility of skin tone on pulse oximetry in critically ill patients: a prospective cohort study — but that study explicitly lists local perfusion, epidermal thickness, and vascular resistance as unmeasured confounders, and the causal work downstream treats the skin–sensor interface as a black box Path-specific effects for pulse-oximetry guided decisions in critical care. We claim that a calibrated measurement of transcutaneous gas flux and of the microbial community at the sensor site, added to existing skin-tone predictors, will (a) explain a measurable share of the residual bias variance and (b) correct a previously unrecognized artifact: part of any gas accumulating under a chamber or clip seated on skin is produced or consumed by the skin microbiome, not by the patient. If melanin metrics turn out to explain everything, that is also a decisive finding — it would close the door on biological confounding with data rather than assumption.

Research Program

The instrument is a miniaturized static flux chamber: a rigid, opaque enclosure of known volume seated on a flat adhesive collar at oximeter-relevant sites (thenar eminence, wrist, earlobe), with in-line non-dispersive infrared sensing of carbon dioxide and electrochemical sensing of oxygen, following the field-portable direct-sensing pattern already proven for soil respiration with the Li-6400-09 chamber Yang et al., 2022). The core relation is the mass-balance equation validated by controlled-release testing, $F = \frac{V}{A}\cdot\frac{dC}{t}$, where $F$ is the surface flux, $V/A$ the chamber's volume-to-area ratio, and $dC/dt$ the accumulation rate Controlled-release testing of the static chamber methodology for direct measurements of methane emissions. Because the accumulation curve bends as headspace concentration rises — the known nonlinearity problem in soil chambers — we fit curvature-aware (exponential) models rather than linear slopes, and keep deployment windows under five minutes to limit microclimate perturbation of the skin.

Three methodological transfer problems get explicit treatment. First, chamber validation: we will benchmark the device against known release rates through a gas-permeable membrane skin phantom, aiming for accuracy comparable to the $\pm 5\%$ achieved by full-scale chambers under optimized configurations Controlled-release testing of the static chamber methodology; no reported result found on Lacuna for any miniature or cutaneous chamber, so this benchmark is new ground. Second, host–microbe attribution: the chamber measures total flux, and the skin's bacterial biofilms consume oxygen and emit carbon dioxide. We separate the two with paired adjacent sites, one pretreated with antiseptic decolonization and one sham-treated, cross-checked against 16S-derived biomass estimates; 16S alone cannot distinguish living from dead cells, so a culture arm (MALDI-TOF on isolates) serves as a viability check, with the caveat — documented head-to-head — that culture severely misses fastidious genera like Corynebacterium that sequencing captures Exploring the interrelationship between the skin microbiome and skin volatiles: A pilot study. Third, measurement ordering: skin-tone colorimetry and spectrophotometry first, chamber flux second (non-destructive), then swab for V4-region 16S sequencing last, so each modality leaves the surface undisturbed for the next.

The cohort is two-stage. Stage one recruits roughly 150 healthy volunteers deliberately stratified and oversampled toward the dark end of the Monk scale, fixing the under-representation at dark tones and extreme hypoxemia that limited the prior prospective study Utility of skin tone on pulse oximetry. Stage two replicates that study's ICU protocol — paired pulse-oximetry and arterial blood gas readings within a strict five-minute window — in about 60 arterial-line patients, adding chamber flux and 16S swabs at the sensor site. Analysis mirrors the prior work's mixed-effects machinery (heart rate, mean arterial pressure, and pH as covariates) with nested model comparison, plus the Random Forest and SHAP attribution pipeline that the skin volatilome study used to combine chemical and microbial data streams Exploring the interrelationship between the skin microbiome and skin volatiles. All data will be released in the OMOP format pioneered by ENCoDE, which showed how prospective skin-tone datasets become reusable community resources ENCoDE – a skin tone and clinical dataset from a prospective trial on acute care patients.

Proposed Experiments

Experiment one validates the chamber against membrane phantoms with known flux, reporting accuracy relative to the $\pm 5\%$ full-scale benchmark. Experiment two tests whether transcutaneous oxygen and carbon dioxide flux, after conditioning on melanin index and perfusion covariates, still varies across skin tone and predicts the oximetry discrepancy $V = \text{SpO}_2 - \text{SaO}_2$; the decisive statistic is a likelihood-ratio test on flux terms added to the skin-tone-only model, and a cross-validated mean absolute error on $V$ compared between models. Experiment three quantifies the microbial respiratory share of the measured flux via the decolonization contrast, and tests whether microbial features (Shannon diversity, Cutibacterium/Staphylococcus/Corynebacterium abundances, copy-number-corrected biomass) predict residual bias. We also report the regulatory metric decomposed: because $A_{\text{RMS}}$ combines bias and variance, devices can pass the FDA's $A_{\text{RMS}} \le 3.0\%$ while being, in the prior study's phrase, "consistently wrong" for dark skin Utility of skin tone on pulse oximetry — we will report directional bias and variance separately by tone decile, which is the honest metric. Optionally, camera-based photoplethysmography at the chamber site can be added using an open biosensing codebase SUMS, for which no reported result found on Lacuna; this is a secondary measurement, not a load-bearing component.

Baselines

The do-nothing reference point is an uncorrected FDA-cleared pulse oximeter judged by aggregate $A_{\text{RMS}}$. This baseline is deceptive rather than weak: the prospective Duke cohort showed it can pass while carrying a -2.40 percentage-point directional bias between Monk-tone extremes Utility of skin tone on pulse oximetry. Beating it means reducing mean directional bias in the darkest tone categories on our cohort until the 95% confidence interval no longer reproduces that -2.40 magnitude, with variance reported alongside rather than pooled into it.

The strongest existing competitor is the skin-tone-only mixed-effects prediction from that same study — the six-variable objective skin model (significant at $p=0.02$) and the Monk Skin Tone scale ($p=0.01$) Utility of skin tone on pulse oximetry. We will run it on our own cohort with the identical devices, anatomical locations, and five-minute pairing window, so the comparison is nested within the same subjects rather than across populations. Our full model adds chamber-derived flux and microbial features to exactly this base. Fairness demands the same tuning budget and the same cross-validation folds for both; the result that counts as beating it is a significant likelihood-ratio test for the added terms together with a materially lower cross-validated error on $V$.

The key ablation isolates the chamber itself: replace chamber-derived flux with a gauze-pad volatile or gas-intensity proxy collected exactly as in the skin volatilome study, which classified body sites at 85–90% accuracy from individual data streams Exploring the interrelationship between the skin microbiome and skin volatiles, and push it through the same prediction pipeline. If the passive proxy predicts bias as well as true kinetic flux, the chamber's contribution collapses to the microbial-artifact correction alone, and we will say so. A second ablation drops the microbial features to test whether the flux-to-bias link survives without the biofilm term.

Why not just…?

Why not just recalibrate devices with skin tone measurements? This is the most serious alternative, and it is partly right: skin-tone metrics already carry significant predictive power ($p=0.01$–$0.02$) Utility of skin tone on pulse oximetry, and a tone-offset correction is far cheaper than any chamber. But that study itself flags local perfusion and epidermal composition as unmeasured confounders, and its darkest-tone and hypoxemic data were thin. Our stage-one experiment is designed precisely to test whether residual bias remains after conditioning on every objective tone metric the prior work used; if the residual is negligible, the cheap recalibration is vindicated by our data, and the chamber's value reduces to correcting the microbial contribution — still a result neither side of that argument currently has.

Why not just use race-based correction offsets? The evidence rules this out directly: self-reported race was not a significant predictor of bias ($p=0.64$) in mixed models where objective skin metrics were Utility of skin tone on pulse oximetry, and retrospective databases like BOLD, with its 49,099 paired readings, record only administrative race categories rather than any optical measurement BOLD: Blood-gas and Oximetry Linked Dataset. A race offset would correct a social label, not the physical attenuator.

Why not just fix the hardware — more wavelengths or radar fusion? Multi-wavelength reflective photoplethysmography Tri-Spectral PPG and RGB-radar fusion CardiacMamba attack motion and illumination artifacts, and optical methods degrade badly in low light Generalization of Video-Based Heart Rate Estimation Methods To Low Illumination and Elevated Heart Rates, with thermal imaging showing inter-device discrepancies above 1°C Impact of Skin Tone, Environmental, and Technical Factors on Thermal Imaging. All of these still model skin as a homogeneous inert medium. Our measurements provide the test hardware-only work lacks: if transcutaneous flux and microbial terms explain residual bias after tone is controlled, no wavelength count fixes that, because the confounder is not in the light path.

Why not just swab with gauze pads instead of building a chamber? Gauze sorption, as deployed over 20-minute windows in the volatilome study, yields relative chemical profiles but no sealed headspace, no known $V/A$, and therefore no way to compute a mass flux Exploring the interrelationship between the skin microbiome and skin volatiles. Our ablation will make this comparison empirical rather than rhetorical.

Why not just culture the microbes instead of sequencing? Culture misses dominant skin taxa — Corynebacterium was robust in 16S data and nearly absent in culture in the same study — so a culture-only design would blind us to the very communities whose metabolism the chamber detects. We keep culture as a viability check, not the profiling tool.

Decisive Checks and Limits

The proposal succeeds if one of two clean outcomes emerges: flux and microbial features add significant predictive power for oximetry bias beyond the strongest skin-tone model, or they add nothing — which, given the confounder list in the existing prospective work, would itself be the first direct evidence that the bias is purely optical and that hardware recalibration by tone is sufficient. Known limits carry over honestly: 16S abundance is relative and primer-biased, relic DNA cannot be fully excluded, and the ICU arm inherits the scarcity of extreme hypoxemia that constrained the Duke cohort. The chamber, once validated to a phantom benchmark, is the piece that is genuinely new: no one has ever measured a calibrated gas flux off human skin at the exact place a medical optical device sits, and that single measurement is what lets the three-way decomposition — melanin, microvasculature, microbiome — be asked at all.