Correct the Sensor, Not the Model: Decoupling Optical from Social Proxy Bias in Pulse Oximetry

When a hospital strips patient race from a clinical prediction system, the model usually recovers it anyway. This is proxy variable bias: high-dimensional features are entangled with protected attributes through physical, behavioral, and historical processes, so removing the label removes nothing. De-Arteaga et al. showed this for text, where scrubbing names and pronouns from biographies leaves gender recoverable from phrasing; Jiang et al. and Loftus et al. showed the same for course histories and postal codes. This proposal is about a setting where the story is more specific, and more fixable, than in any of those examples: in the intensive care unit, one of the leakiest features a blinded model ingests is the pulse oximeter reading itself. Epidermal melanin absorbs the sensor's light, so the device systematically over-reports blood oxygen in darker-skinned patients — a hardware artifact, not a statistical pattern in a spreadsheet. The protected attribute does not merely correlate with the input; it physically causes the input to be wrong.

That distinction changes what debiasing should mean. The dominant mitigation for proxy leakage — adversarially forcing a representation to be invariant to the sensitive attribute, as in Zhang et al. — is built for settings where the proxy link is spurious. Here it is causal: melanin is the very thing that determines how the reading must be corrected, so stripping melanin information from an encoder discards the correction signal. The consequences of not correcting are severe and quantified: darker-skinned patients are "consistently wrong" rather than noisy, with a mean directional bias of -2.40% (95% CI -4.32 to -0.48) between the lightest and darkest Monk Skin Tone strata in Hao et al., producing hidden hypoxemia — a reassuring $SpO_2$ reading over a dangerously low true saturation — and, per Zhang et al., mechanical ventilation that runs 2.3 hours longer (eICU) and 7.4 hours longer (MIMIC-IV) for Black patients via this measurement error alone. Yet the same error can pass the FDA clearance threshold $A_{RMS} \le 3.0\%$, because darker-skinned cohorts have lower error variance alongside their higher mean bias.

Core Claim

The claim has two parts. First, conceptual: in optical physiological sensing, melanin-induced attenuation is not a proxy to be scrubbed but a measurement mechanism to be inverted, and the right tool is an explicit, physics-parameterized error model, not representation-level invariance. Second, empirical: a hierarchical measurement model, fitted to paired non-invasive readings and arterial blood gas truths without any demographic input, can (a) recover a latent per-patient attenuation state that tracks objective skin tone better than self-reported race does, (b) cut the hidden-hypoxemia gap and the bias-mediated ventilation disparity, and (c) cleanly separate the portion of clinical disparity that flows through the optical channel from the portion that flows through social and systemic channels — which no existing work does.

The need for this decoupling is a live contradiction in the literature. Hao et al. found that self-reported race has no statistically significant power to predict oximetry error once objective skin tone is in the model ($p = 0.64$ for race versus $p = 0.01$ for the Monk scale), yet Zhang et al. root their causal mediation analysis in exactly that self-reported race node. Their $V$-specific Direct Effect conflates "bias caused by melanin attenuating photons" with "disparity caused by being racialized in an ICU." Until the two channels are separated, debiasing algorithms may adjust sensors using social labels that do not predict the error, and policy may blame hardware for disparities that are institutional.

What the Physics Buys

The biophysical literature supplies the functional form for free. Putcha et al., in a controlled live porcine model with naturally pigmented and non-pigmented skin patches, showed that melanin does not add a constant offset: it compresses the slope of the red/NIR modulation ratio against true saturation, suppressing the red pulsatile component disproportionately exactly when saturation falls. So the bias is state-dependent — it grows in the hypoxemic region where patients are sickest and data is sparsest. This is what defeats constant-offset corrections, but it is also an opportunity: a distortion that varies systematically with the true state leaves a distinctive fingerprint in each patient's record.

Concretely, we model each paired measurement as

$$SpO_2^{\text{obs}}_{ij} = h\big(S_{ij};\, \theta_i\big) + \varepsilon_{ij},$$

where $S_{ij}$ is the true arterial saturation for patient $i$ at time $j$ (observed only at blood-gas draws), $h$ is a monotone, saturation-dependent distortion drawn from a low-dimensional family whose shape is fixed by the porcine biophysics (slope compression that strengthens as $S$ drops below 88%), $\theta_i$ is a scalar latent attenuation state standing in for effective melanin density at the sensor site, and $\varepsilon$ is zero-mean sensor noise. Because a patient's paired observations trace out the graph of $h$ over a range of saturations, $\theta_i$ is identifiable from repeated pairs alone — no demographic label is needed at any point. The physics family is what makes this work where a free per-patient fit would not: with few parameters shared across patients, the model pools strength, so the sparse, clinically critical hypoxemic region is covered by the shared family rather than by per-patient extrapolation. Correction is then simply inverting $h$ under the posterior over $\theta_i$.

This design deliberately sidesteps two known failures of the causal-fairness toolkit. Kusner et al. note that counterfactual guarantees hinge on an unfalsifiable causal DAG, and that enforcing them costs accuracy (test RMSE on the law-school benchmark rises from 0.873 to 0.918–0.929). Our model makes no counterfactual claim and removes no predictive information — it corrects a measurement, so accuracy for darker-skinned patients should improve, with no fairness–accuracy trade-off. Shpitser et al. supply the mediation machinery we reuse for evaluation, and Yfantidou et al. supply the motivation: because waveform-carried bias is invisible to human auditors, blinded deployment in wearables is, in their words, exceptionally dangerous — our approach keeps the correction auditable and demographic-free at inference time.

Research Program

Three components. First, fit the hierarchical measurement model on BOLD, the harmonized MIMIC-III/IV/eICU benchmark of 49,099 time-aligned $SpO_2$–$SaO_2$ pairs (70–100% saturation, roughly 25% minority representation), using self-reported race only for post-hoc evaluation, never as a model input. Second, validate the latent attenuation state $\theta$ against ground truth on ENCoDE, the prospective Duke cohort of 128 acute-care patients with eight objective skin-tone metrics (Monk scale, Fitzpatrick, melanin index, colorimetry) across 16 body sites, linked to 521 paired measurements. The test: does inferred $\theta$ predict Monk Skin Tone and melanin index at least as strongly as self-reported race predicts error? Hao et al.'s finding — skin tone $p = 0.01$, race $p = 0.64$ — is the pattern we should reproduce with an unlabeled latent. Third, run the downstream proxy-leak test: train blinded clinical models on MIMIC-IV (ventilation duration, mortality, extubation readiness) with either raw or corrected $SpO_2$, probe how much race an adversary can recover from the inputs and residuals (the audit style of De-Arteaga et al. and Jiang et al.), and recompute the $V$-specific Direct Effect of Zhang et al. with their self-normalized doubly robust estimator, which achieved an error of 0.003 against 0.024 for an X-learner in their semi-synthetic benchmarks — a validated tool we adopt unchanged.

Baselines

The "do nothing" reference point is the FDA-cleared device output used as-is. On it, Hao et al. report the -2.40% Monk-stratum bias and the paradox that these same devices can satisfy $A_{RMS} \le 3.0\%$; on BOLD we will compute the per-race hidden-hypoxemia rate ($SpO_2 \ge 88\%$ while $SaO_2 < 88\%$) as the headline number. BOLD itself reports no numerical RMSE splits by demographic group — no reported result found on Lacuna for BOLD's per-group agreement figures — so we regenerate them. Beating "do nothing" means closing the between-group hidden-hypoxemia gap and bringing per-group mean directional bias to statistical non-significance, while keeping or improving overall $A_{RMS}$.

The strongest existing corrective competitor is per-race affine recalibration: fit a separate linear (or constant-offset) mapping from $SpO_2$ to $SaO_2$ for each self-reported race group on BOLD, then apply it at inference. This is the strongest because it is the natural response once one accepts the directional-bias literature, it uses the same paired data, and it subsumes the crude "+2.4% for Black patients" suggestion. It will be run under identical conditions: same train/test splits, same cross-validated tuning budget, same compute, evaluated on the same held-out strata. We beat it if we achieve lower mean directional bias and lower hidden-hypoxemia disparity on held-out ENCoDE skin-tone strata — where within-race tone variation exists and a race-keyed correction cannot follow it — and better calibration in the $SaO_2 < 88\%$ region, where Putcha et al. show the distortion is non-linear. A second, analytic baseline is not a corrector but the evaluation machinery of Zhang et al.: their VDE estimates on raw data (2.3 hours eICU, 7.4 hours MIMIC-IV for ventilation duration; -0.21 and -0.46 percentage points for initiation) are the pre-correction benchmark our corrected readings must shrink.

The key ablation isolates the physics. We retrain the identical hierarchical model with $h$ restricted to a per-patient constant shift (no saturation dependence, no shared slope-compression family). This keeps the label-free latent inference and the hierarchy, removing only the state-dependent distortion shape. If the full model wins only in the hypoxemic region and on ENCoDE tone prediction, the physics parameterization — not the hierarchy — is doing the work, which is precisely the claim. A second ablation replaces the inferred $\theta_i$ with observed race, testing whether label-free inference matches label-based correction.

Why not just…?

Why not just blind the model to race and skin tone? This is fairness through unawareness, and the evidence against it is unanimous across modalities: De-Arteaga et al., Jiang et al., and Loftus et al. all show models reconstruct the omitted attribute from remaining features, and Kusner et al. show formally that unaware predictors still shift under counterfactual attribute changes because causal descendants remain in the input. Deng et al. document that practitioners keep choosing this route anyway, for want of guardrails. Our leak experiment measures it directly in this setting, and Yfantidou et al. explain why it is worse here than elsewhere: the distortion is physically baked into the reading, so blinding conceals the bias while leaving the wrong number in the chart.

Why not just add a fixed race-based offset? Three cited findings say it falls short. Putcha et al. show the error is a slope change, not an offset, so a constant correction miscalibrates precisely in severe hypoxemia. Hao et al. show race does not predict the error ($p = 0.64$) while individual tone does ($p = 0.01$), so a race-keyed offset misclassifies dark-skinned patients within lighter categories. And Charpignon et al. document the clinical history of hardcoded race corrections normalizing worse care. Honestly, though, this is the one alternative that could nearly match us on BOLD alone, because BOLD carries only race labels; if per-race affine calibration matches our model on BOLD's aggregates, we will report that and let the ENCoDE within-race strata, the hypoxemic region, and deployment without demographic input decide the rest. That comparison is built into the experimental design rather than argued around.

Why not just adversarially train a skin-tone-invariant encoder on the waveforms? This is the transfer of Zhang et al. to PPG, and it faces a data problem before a conceptual one: no benchmark in the literature has both raw waveforms and skin-tone labels. PaPaGei pre-trains on 57,000+ hours of PPG and beats generic foundation models by an average 6.3% AUROC across 20 tasks, but its corpora (MIMIC, VitalDB, MESA) lack standardized skin-tone representation, and no reported result found on Lacuna for PaPaGei's disaggregated Fitzpatrick subgroup performance. ENCoDE has tone labels but only 128 patients and no raw waveforms. Conceptually, invariance is the wrong target here: melanin causes the distortion, so an encoder that cannot represent skin tone cannot compensate for it — the fairness–accuracy cost that Zhang et al. themselves report (accuracy 86.0% falling to 84.5% on UCI Adult while equalizing false positive rates from 0.0248/0.0917 to roughly 0.065/0.070) would, in this setting, be paid in exactly the patients whose readings need correcting. Still, if paired waveform-plus-tone data emerges, this becomes a real competing approach, and our measurement model can serve as its evaluation floor.

Proposed Experiments and Decisive Checks

Experiment one (measurement correction): fit on BOLD, evaluate per-race mean bias, per-race $A_{RMS}$, and hidden-hypoxemia rates before and after correction, against raw readings, race-based affine recalibration, and the constant-shift ablation. The decisive check is the $SaO_2 < 88\%$ slice, where the physics family must hold and offsets must fail; ENCoDE's authors note their own cohort lacked severely hypoxemic patients, so this check runs on BOLD, whose range spans 70–100%.

Experiment two (latent validation): on ENCoDE, regress the inferred $\theta_i$ against Monk Skin Tone, melanin index, and self-reported race. The decisive check is the Hao pattern inverted — the unlabeled latent should track tone significantly and race weakly. If it tracks race instead, the model has learned a social proxy, not an optical state, and the core claim fails.

Experiment three (downstream decoupling): retrain blinded ventilation-duration and mortality models on MIMIC-IV/eICU with corrected $SpO_2$, probe race reconstructability, and recompute the VDE with the estimator of Zhang et al.. The decisive check is bidirectional and honest: if the bias-mediated ventilation effect (2.3 and 7.4 hours) collapses toward zero, the optical channel was real and is now closed; if it barely moves, the disparity is substantially social rather than optical — a finding that redirects the intervention and contradicts the hardware-centric reading of the current literature. Either outcome is informative, and the regulatory check closes the loop: corrected readings must simultaneously pass $A_{RMS} \le 3.0\%$ and eliminate per-group mean directional bias, resolving the paradox Hao et al. exposed in the clearance standard itself.