Retrieve the Veil First: Degeneracy-Aware Optical Oximetry That Transfers Across Centers

Pulse oximeters estimate blood oxygen by comparing how much red and near-infrared light the pulsing blood absorbs. Skin pigment (melanin) sits on top of the blood like a tinted filter and absorbs red light far more than near-infrared. As a result, devices overestimate oxygen in patients with darker skin, and they do so most when the patient is truly hypoxic. Planetary scientists face the same structure when they look at Titan or Mars through a hazy atmosphere. Their answer is not to learn a correction from labeled examples. They model the haze physically, estimate it from wavelengths where the haze dominates the signal, and then recover the surface from the remaining "windows." We propose doing the same for camera and reflectance oximetry. A fast learned simulator of light moving through layered skin is inverted probabilistically. Color channels that are mostly sensitive to the shallow, melanin-rich layer estimate the "veil," and the deep-penetrating channels then recover oxygen saturation. The output is a full distribution over oxygen, not a single number, so the device can say "I cannot tell" when the physics is genuinely ambiguous.

This matters for two reasons beyond fairness. Rapid optical screening, for example with a smartphone camera, is exactly where hardware is uncalibrated and cohorts are tiny and single-site. Models learned from such data inherit one center's skin tones, altitude, and device quirks. An inverse model trained on simulation and conditioned only on a device's measured spectral response needs no patient labels from a new site. That turns the single-center problem from a data-collection problem into a calibration-file problem.

Core Claim

Two-wavelength oximetry cannot separate "dark skin, normal oxygen" from "light skin, low oxygen" from a single reading. This ambiguity is a property of the measurement, not of the model, so no amount of training data fixes it. Adding one shallow-penetrating channel and inverting a physical forward model removes most of the pigmentation dependence of the error. Where ambiguity remains, a posterior that can represent several distinct solutions flags it instead of averaging the solutions into a confident wrong answer. Our testable claims are three. First, on held-out centers, this retrieval matches or beats per-skin-tone recalibration even when that recalibration is given the true pigmentation group. Second, its posterior width predicts its own errors in the hypoxic, pigmented regime. Third, its physically predicted bias shape (bias that grows as true saturation falls) appears in multi-center intensive care unit (ICU) records.

Research Program

The forward model maps tissue parameters to what a given device sees. We write $\theta$ for the tissue state: arterial saturation $S$, epidermal melanin concentration $m$, epidermal thickness $d$, dermal blood volume, and scattering. For each device channel $c$ with measured spectral sensitivity $s_c(\lambda)$, the predicted pulsatile-to-baseline ratio is

$$y_c(\theta) = \int s_c(\lambda)\, f(\lambda; \theta)\, d\lambda ,$$

where $f(\lambda;\theta)$ is the wavelength-resolved pulsatile reflectance of layered skin. Computing $f$ by Monte Carlo photon simulation is far too slow for bedside use. We therefore reuse the method of Learning to simulate realistic human diffuse reflectance spectra. That work trained a neural network on tens of millions of multilayer Monte Carlo simulations and reproduced their output 134,000 to 246,000 times faster. It is a forward engine only, and its authors do not build the inverse. We extend it in two ways. We simulate the pulsatile component, by differencing simulations at systolic and diastolic blood volume. We also sample melanin over the full human range together with deep hypoxia, a combination that dataset was not designed around.

The inverse is an amortized posterior network: a conditional density estimator trained on simulated pairs to output $p(\theta \mid y, s)$. It is conditioned on the device response $s$, so one network serves many devices. Two design choices come from planetary and remote-sensing retrievals. The first concerns the shape of the posterior. Inference over radiative transfer models using variational and expectation maximization methods found on a vegetation-canopy benchmark that a Gaussian variational posterior reached Kullback–Leibler divergence 0.0208 against the true posterior, while multimodal Monte Carlo EM reached 0.0123 at much higher cost. The Gaussian version also averaged distinct solutions. We therefore use a flow or mixture posterior and test the Gaussian version as an ablation. This paper constrains the design: it rules out unimodal amortized inference. The second choice is input normalization. Learned spectral decoloring enables photoacoustic oximetry inverted wavelength-dependent light corruption in photoacoustic imaging with a per-pixel network trained on simulation. It used sum-to-one spectral normalization to remove dependence on absolute light intensity, and it recovered a 5–95% saturation range where linear unmixing bottomed out near 40%. We adopt that normalization and its sim-to-real strategy. We differ in three ways: we target reflectance PPG (photoplethysmography, the pulsing light signal used by oximeters and cameras) rather than photoacoustics, we output a distribution rather than a point estimate, and we vary pigmentation systematically. The original validation used only a few healthy volunteers.

The "veil channel" is the mechanism that carries over from Titan. Detailed chemical composition analysis of the Soi crater region on Titan retrieved surface albedo through haze using several atmospheric windows. That paper also issues a warning: its models failed at the shortest windows (0.94 and 1.08 µm) because laboratory optical constants were missing. We expect the same failure mode in the green and blue camera channels, where melanin and keratin optical constants and scattering in calluses are least certain. The program therefore includes an explicit check. We inflate the noise model for the short channels and test whether the veil estimate survives. If it does not, the claim reduces to "the posterior correctly reports ambiguity," which is still useful but weaker.

Finally, we analyze identifiability directly. For each channel configuration (red and near-infrared only; camera RGB; RGB plus near-infrared), we compute where in $(S, m)$ space the posterior becomes broad or bimodal. This yields a map of where each device class is guaranteed to be ambiguous, independent of any learned model.

Baselines

The naive reference is the standard ratio-of-ratios estimate. It computes the modulation ratio from red and near-infrared (or red and green for cameras), with a single calibration curve fit on the pooled training cohort and no pigmentation information. It defines the uncorrected bias we aim to remove.

The strongest simple competitor is per-pigmentation recalibration. It fits a separate slope and intercept of saturation against modulation ratio for each skin-tone stratum. Characterizing the influence of skin pigmentation on pulse oximetry showed in pigs that melanin changes the slope of this relation, not just its offset, so a slope-aware recalibration is exactly what that evidence recommends. We run it in two forms: with skin tone estimated from the device, and with oracle pigmentation labels. It is fit on the same training centers as our method and evaluated on the same held-out center. Beating the oracle form out of center is the bar for claiming physics adds something beyond knowing skin tone.

The strongest learned baseline is the camera network of Smartphone camera oximetry in an induced hypoxemia study. That paper reports mean absolute error 5.00% and root-mean-square accuracy $A_{rms}$ 5.55% over 70–100% saturation, with 81% sensitivity and 79% specificity for saturation below 90%. Its six-subject evaluation was not leave-one-center-out. We retrain its architecture on our training centers with the same tuning budget as our method. Beating it means lower $A_{rms}$ on the held-out center, ideally under the 3.5% clearance threshold that paper cites, together with higher hypoxemia sensitivity at matched specificity.

We also include the multi-site, multi-task fusion model from Summit Vitals via its released code (thuhci/SUMS). That paper reports only relative improvements (7.6–10.6% lower mean absolute error from face-plus-finger fusion, 17.8% from multi-task training). No absolute error figure is reported on Lacuna for SUMS, so we will measure it ourselves.

The ablations isolate the key ingredients. The first is an LSD-style point-estimate network trained on the same simulations: same physics, no posterior. The second is our method with a Gaussian posterior. The third is our method with the short-wavelength veil channel removed.

Why not just…?

Why not just recalibrate slope and intercept per skin-tone group? This is the real threat, and the comparison with oracle pigmentation labels is designed to settle it. Our expectation that it falls short rests on two points. Recalibration needs labeled hypoxic data across skin tones at every site and device, which is the scarce resource in a fragmented system. And a coarse skin-tone category does not capture continuous melanin, epidermal thickness, or calluses, the last of which the smartphone study found dampened the signal. If oracle recalibration matches our method on held-out centers, we will report that physics adds nothing beyond a melanin measurement. We would then reposition the work as a label-free melanin estimator that feeds the simple recalibration.

Why not just train a bigger black-box network on more diverse data? The identifiability analysis is the answer. Where two tissue states produce the same readings, any regressor must output a compromise, and more data cannot fix that. The only fix is more informative channels or an honest statement of uncertainty. The retrained smartphone network will show whether this matters empirically, especially below 85% saturation, where that paper reports its accuracy degrades. Diverse labeled hypoxia data across centers is also exactly what fragmentation and governance make scarce, as argued in The role of machine learning in clinical research.

Why not just add wavelengths and do linear unmixing? Linear unmixing assumes every wavelength reaches the blood equally. Learned spectral decoloring showed this assumption compresses estimates and misses deep hypoxia. Measuring and imaging of transcutaneous bilirubin, hemoglobin, and melanin needed Monte Carlo-derived regression corrections even for a contact probe. We include multi-channel linear unmixing on the camera data as an additional reference point.

Why not run Monte Carlo inversion directly? It is accurate but costs roughly 130 GPU-weeks at dataset scale, per Learning to simulate realistic human diffuse reflectance spectra. We use it only offline, to generate training data and to validate the surrogate on a held-out set of simulations.

Proposed Experiments

The first test is in simulation. We measure posterior coverage, and error stratified by $(S, m)$, for each channel configuration. The decisive check is whether the regions the identifiability analysis marks as ambiguous are where point estimators fail and our posteriors widen.

The second test is real-data transfer under a leave-one-center-out protocol across the smartphone hypoxemia frames (camera RGB, co-oximetry reference, 70–100%), the SUMS high-altitude videos (RGB, reference oximeter), and the porcine pigmented-skin red/near-infrared recordings with blood-gas reference. Access to the first and third datasets must be confirmed with their authors. That is a real risk, and if access fails we would run a small induced-hypoxemia collection spanning skin types. The porcine data has only two wavelengths, so we cannot test the veil channel there. Its role is to test a prediction: posterior width should grow specifically in pigmented skin at low saturation, mirroring the slope loss that study measured (that paper's file summary reports no slope coefficients). The metrics are mean absolute error, $A_{rms}$, bias as a function of true saturation within each pigmentation stratum, sensitivity and specificity below 90%, and the rate of occult hypoxemia among confident predictions.

The third test is a falsification check at ICU scale. eICU (about 37,222 admissions across many hospitals) and MIMIC-IV (about 4,897) do not contain raw multi-wavelength signals, so we cannot correct their readings. We can, however, test the shape the physics predicts. Bias in SpO₂ (the oximeter reading) against SaO₂ (arterial saturation from blood gas) should grow as SaO₂ falls for Black patients, rather than staying a constant offset. Self-reported race is only a crude proxy for melanin, which limits how cleanly this test isolates pigmentation. Blood gases are drawn when clinicians are worried, so we weight time-matched pairs for informative sampling, following the argument of Nonstationary Gaussian Process Regression for Evaluating Repeated Clinical Laboratory Tests. We also compare race-driven variance with between-hospital variance in eICU. Path-specific effects for pulse-oximetry guided decisions in critical care linked oximeter bias to 2.3 hours (eICU) and 7.4 hours (MIMIC-IV) longer ventilation for Black patients but treated the device as a black box. Our check asks whether that bias has the physical form our retrieval removes.

The project fails if two things both hold. First, oracle per-group recalibration matches our method on held-out centers. Second, posterior width does not predict error in the hypoxic, pigmented regime.