Consistently Wrong Sensors, Compounding Policies: Melanin Bias in Closed-Loop Ventilation Control

Pulse oximeters clip onto a finger and estimate blood oxygen ($SpO_2$) from light absorption. On patients with darker skin, melanin absorbs light before it ever reaches the blood, and the result is not random noise: the device reads systematically high, and it does so with low variance — it is, as one prospective ICU study put it, more consistently wrong. A patient can be in hidden hypoxemia, with true arterial saturation ($SaO_2$) below 88%, while the monitor reports a reassuring 92% or more. Clinicians acting on that reading keep Black patients on mechanical ventilation 2.3 to 7.4 hours longer than White patients, an effect a recent causal study traced specifically to the oximeter error itself Path-specific effects for pulse-oximetry guided decisions in critical care.

Meanwhile, reinforcement learning agents are being trained to control ventilation settings themselves. DeepVent, an offline RL agent trained on 19,780 ventilated MIMIC-III patients, learns to set tidal volume, PEEP, and inspired oxygen, and under fitted-Q evaluation reaches a mean initial state value of 0.743 versus 0.502 for the recorded physician policy Towards Safe Mechanical Ventilation Treatment Using Deep Offline Reinforcement Learning. But DeepVent, like every ventilation RL system we found, treats $SpO_2$ as a pristine state variable. This proposal asks what happens when a sequential decision agent closes the loop on a sensor that is directionally biased by skin pigmentation, and how to build an agent that does not compound that bias into multi-hour clinical disparities.

Core Claim

The claim is that melanin-induced oximetry error is a closed-loop amplifier, not a calibration nuisance. Three facts from the literature, none of which have been combined, imply this. First, the error is directional and low-variance rather than zero-mean: across 521 paired readings, the difference in bias between the lightest and darkest Monk Skin Tone (MST) categories was -2.40% (95% CI -4.32% to -0.48%), worsening precisely in the desaturated range where decisions matter most Utility of skin tone on pulse oximetry in critically ill patients: a prospective cohort study. Second, in sequential decision-making, systematic state perturbation diverges state-visitation distributions across groups — the decomposition theorem of Towards Return Parity in Markov Decision Processes splits return disparity into reward, policy, and state-visitation terms, and a persistent input bias feeds the third term multiplicatively over the horizon. Third, static one-shot corrections demonstrably fail in dynamic systems: in the 5-step lending simulation of Achieving Long-Term Fairness in Sequential Decision Making, a static fairness-constrained model saw long-term unfairness escalate to 0.481 while dynamically retrained models held it at 0.002.

We therefore claim two things. Empirically: a DeepVent-class agent run on MST-skewed $SpO_2$ will produce ventilation-duration disparities that substantially exceed what its per-step bias suggests, because the agent's weaning policy is itself a function of the biased state — the bias writes itself into the policy and then compounds. Constructively: an agent that maintains a belief over true saturation, conditioned on continuous skin pigmentation and calibrated with the sporadic arterial blood gases that do exist, can largely remove this compounding without sacrificing value.

The closest existing work is the path-specific-effects study above, which found the 2.3–7.4 hour disparity under human clinicians, but explicitly models static, stay-aggregated effects, uses binary race rather than any pigmentation measure, and deploys no policy Path-specific effects for pulse-oximetry guided decisions in critical care. We differ on all three axes: dynamic closed-loop simulation with a trained policy, continuous MST-parameterized error, and a mitigation agent. DeepVent supplies the RL machinery but is totally blind to measurement error Towards Safe Mechanical Ventilation Treatment Using Deep Offline Reinforcement Learning; we reuse its pipeline and corrupt its observation channel.

Research Program

The program has three components: a pigmentation-parameterized measurement-error model, a closed-loop "bias-in-the-loop" simulator, and a bias-aware control agent.

The error model. We learn $O_t = S_t + \delta(S_t, m) + \varepsilon_t$, where $O_t$ is the observed $SpO_2$, $S_t$ the true saturation, $\delta$ a directional error function of true saturation and a continuous pigmentation variable $m$, and $\varepsilon_t$ small noise. The pairing data comes from BOLD: Blood-gas and Oximetry Linked Dataset, which harmonizes MIMIC-III, MIMIC-IV, and eICU into 49,099 paired $SpO_2$–$SaO_2$ measurements with roughly 25% minority representation. The obstacle is that BOLD and the MIMIC databases carry race labels but not skin tone, and race is a poor predictor of oximetry error (p=0.64 in the Duke cohort, versus p=0.01 for MST Utility of skin tone on pulse oximetry in critically ill patients). We therefore treat MST as a latent variable: ENCoDE – a skin tone and clinical dataset from a prospective trial on acute care patients provides 167 skin-tone features (MST, Fitzpatrick, spectrophotometry) across 16 body locations on 128 acute-care patients, from which we learn a race-and-covariate-conditional MST distribution and marginalize over it. This is a deliberate move beyond the categorical sensitive attributes used by every dynamic-fairness framework we found, including the counterfactual distribution-mapping approach of A Distribution Mapping Approach to Counterfactually Fair Reinforcement Learning. We will be honest about the extrapolation risk: neither the Duke cohort nor ENCoDE has adequate coverage of severe hypoxemia ($SaO_2 < 88\%$) or the darkest MST extremes, so $\delta$ in the most dangerous regime is identified only weakly, and we will carry that uncertainty through as an error model ensemble rather than a point estimate.

The simulator. We re-train a DeepVent-style conservative Q-learning agent on the MIMIC-III ventilated cohort, then wrap it in a closed-loop patient simulator whose true saturation dynamics are fit from BOLD pairs. At each control step the agent sees $O_t$ through the error model, not $S_t$. The simulator's validity check is demanding: run the clinician policy (or a behavior-cloned stand-in, FQE value 0.572 in DeepVent) through the biased observation channel and check that the resulting ventilation-duration disparity lands in the neighborhood of the real-world causal estimates — 2.3 hours on eICU and 7.4 hours on MIMIC-IV Path-specific effects.... If the simulator cannot reproduce that magnitude, we fix it before trusting anything downstream. The simulator also lets us reconcile a genuine contradiction in the record: a raw demographic audit of MIMIC-IV found Black and Hispanic patients had shorter average ventilation durations MIMIC-IF: Interpretability and Fairness Evaluation of Deep Learning Models on MIMIC-IV Dataset, while the causal study found the oximeter-mediated component made them longer. Our decomposition, applying the return-parity machinery of Towards Return Parity in Markov Decision Processes to the simulator, can separate access-related reward disparities (which push one way) from measurement-mediated state-visitation divergence (which pushes the other), something neither paper could do.

The agent. The mitigation is a belief-state controller: true saturation is latent, the agent filters $O_{1:t}$ given $m$ into a posterior over $S_t$, and acts on the posterior. Two refinements matter. Arterial blood gases are drawn selectively — a clinician orders one when something looks wrong, which is exactly what melanin bias hides — so the gold-standard labels are missing precisely where they are needed, the selective-labels problem formalized by Long-term Fairness with Selective Labels. We adapt its machinery (an imputation predictor plus Rényi divergence regularization between the drawn and not-drawn populations, which reduced true disparity to 0.03 versus 0.30–0.40 for standard fair-RL baselines on the PowerED study) to the blood-gas draw decision. And, following Remembering to Be Fair: Non-Markovian Fairness in Sequential Decision Making, the filter keeps a temporal history, since the fairness-relevant quantity (cumulative hidden-hypoxemia exposure) is not a function of the current state alone.

Baselines

The naive reference point is doing nothing: the DeepVent-style conservative Q-learning agent trained and deployed on raw, uncorrected $SpO_2$ Towards Safe Mechanical Ventilation Treatment Using Deep Offline Reinforcement Learning. This is the strongest existing method on this question — it is the current best-performing offline RL system for ventilation control, reporting a fitted-Q-evaluation mean initial state value of 0.743 against 0.502 for the physician policy and 0.572 for behavior cloning on its 19,780-patient MIMIC-III cohort — and it is the system whose blindness to measurement error defines the gap. It will be run identically to our agent: same cohort, same 36-variable state space, same CQL hyperparameter search budget and compute, with only the observation channel differing.

The strongest mitigating competitor is a static additive offset: subtract a fixed correction from $SpO_2$ for darker-skinned patients before it enters the agent, using the published lightest-versus-darkest difference of -2.40% as the offset Utility of skin tone on pulse oximetry in critically ill patients. It gets the same simulator, the same tuning budget for its one parameter, and the same evaluation. Beating it means reducing simulated ventilation-duration disparity and hidden-hypoxemia exposure time while the offset either fails to close the gap or, as we predict from the nonlinearity of $\delta$, trades hidden hypoxemia for false hypoxemia alarms in well-oxygenated regimes. A second competitor conditions the correction on binary race instead of continuous MST; beating it isolates the value of the pigmentation variable itself, given that race explained oximetry error at p=0.64 versus p=0.01 for MST in the Duke cohort. The key ablation replaces the directional error $\delta$ with zero-mean Gaussian noise of matched RMS magnitude while holding everything else fixed; if our compounding claim is right, the directional version must produce substantially larger disparities, and if it does not, the core mechanism claim is falsified. Finally, the recorded physician policy serves as an external anchor: the uncorrected agent's disparity should be at least as large as the real-world 2.3–7.4 hour figures, and our corrected agent should fall materially below them while keeping its fitted-Q value above the physician policy's 0.502 — accepting that some value loss may be unavoidable, as the fairness literature consistently shows a cost (accuracy fell from roughly 0.82 to 0.68 in the constrained model of Achieving Long-Term Fairness in Sequential Decision Making).

Why not just…?

Why not just recalibrate the oximeter or apply a static correction? Because the bias is state-dependent, not constant. The Duke study found the -2.40% lightest-to-darkest difference with a confidence interval of -4.32% to -0.48%, and the error characteristically worsens during acute desaturation Utility of skin tone on pulse oximetry in critically ill patients — exactly the regime a weaning agent probes. A fixed offset calibrated at moderate saturations under-corrects when the patient is crashing and over-corrects when they are fine, converting one clinical error into another. More fundamentally, the lending simulations of Achieving Long-Term Fairness in Sequential Decision Making show static corrections degrade over sequential interaction (long-term unfairness 0.481 for a static fair model versus 0.002 for dynamic retraining). Still, this is a real threat: if the simple offset closes, say, 90% of the disparity gap in simulation at negligible value cost, the elaborate machinery is not worth it, and our experiments are designed so that outcome would be visible and reportable.

Why not just drop $SpO_2$ and use arterial blood gases? Because blood gases are invasive, risky, and sporadic, while ventilation control needs continuous adjustment of PEEP and inspired oxygen Towards Safe Mechanical Ventilation Treatment Using Deep Offline Reinforcement Learning; BOLD exists precisely because the two are sampled at wildly different rates. Removing $SpO_2$ turns the problem into a severely partially observed process where value estimation variance explodes. Moreover, the decision to draw a blood gas is itself biased by the corrupted $SpO_2$ reading — the selective-labels failure mode that Long-term Fairness with Selective Labels proves offers zero population-level guarantees when you optimize only on observed labels. Our approach faces this head-on rather than assuming the sparse labels are ignorable.

Why not just use a bigger generic time-series model to learn the bias implicitly? The evidence says capacity does not fix this. Generic time-series foundation models underperformed a much smaller domain-specific model on physiological signals, and even the domain-specific PaPaGei: Open Foundation Models for Optical Physiological Signals — which improved classification AUROC by 6.3% over generic backbones and explicitly audits across Fitzpatrick tones — still showed persistent skin-tone performance disparities. Likewise, in MIMIC-IF, an IMV-LSTM achieving 0.955 AUROC on mortality still exhibited severe subgroup disparity. Implicit learning cannot recover a signal the sensor never emitted; the bias is information-destroying at the observation level, and only an explicit model of $\delta$ plus the sparse ground-truth labels can invert it. We will nonetheless run a capacity-increased uncorrected agent as part of the do-nothing baseline family to test this empirically rather than by assertion.

Why not just run the existing path-specific causal analysis? That work is the closest competitor and it already quantified the harm Path-specific effects for pulse-oximetry guided decisions in critical care. But it is static, stay-aggregated, uses binary race, and evaluates no policy — its own stated limitations. It tells us clinicians' past decisions were skewed; it cannot tell us what an autonomous controller will do, whether the bias compounds or saturates, or whether any mitigation works. Our proposal is the dynamic, policy-level, MST-continuous continuation that paper explicitly leaves open, and its 2.3–7.4 hour figures become our simulator's calibration targets rather than our result.

Proposed Experiments and Decisive Checks

Experiment one measures compounding. On the MIMIC-III-based simulator, we compare four agents differing only in the observation channel: true $S_t$ (oracle), uncorrected $O_t$, zero-mean noise of matched RMS, and MST-directional bias. The prediction, derived from the return-parity decomposition — directional input bias grows the state-visitation divergence term over the horizon rather than averaging out — is a monotone ordering: oracle smallest, matched-noise near oracle, directional bias largest, with the directional-to-matched-noise gap growing in the horizon length. If matched noise produces comparable disparities, the core claim is wrong and the "consistently wrong" characterization carries no dynamic penalty; we would report that and stop.

Experiment two tests mitigation. The belief-state agent with selective-label-aware blood-gas calibration runs against the uncorrected agent, the static offset, and the race-conditioned corrector on the same simulator and the same tuning budget. Primary metrics are simulated ventilation-duration disparity between the top and bottom MST terciles (anchored to the real-world 2.3–7.4 hour causal figures), cumulative hidden-hypoxemia exposure time ($SaO_2 < 88\%$ while $SpO_2 \ge 92\%$), and fitted-Q value (which must stay above the physician policy's 0.502, with the gap to 0.743 reported as the cost of robustness). Success is cutting the duration disparity by at least half relative to uncorrected control while keeping hidden-hypoxemia exposure below the static-offset agent, which we predict will over-alarm in the well-oxygenated regime.

Experiment three confronts the identification weakness head-on. Because the error model is only weakly identified in severe hypoxemia and at dark MST extremes, we re-run experiments one and two across the full ensemble of plausible $\delta$ functions, reporting the disparity range rather than a point estimate. We also cross-check the observation model on BOLD's 49,099 held-out pairs, and, as an external sanity check on the selective-label machinery, verify that our imputation of un-drawn blood gases reduces the drawn-versus-not-drawn divergence in the manner Long-term Fairness with Selective Labels achieved on PowerED. A final decisive check mirrors the regulatory failure the Duke study identified — that devices pass the FDA's $A_{RMS} \le 3.0\%$ standard while masking severe bias because darker-skinned cohorts show lower variance with higher mean shift Utility of skin tone on pulse oximetry in critically ill patients: we will show whether an agent selected on RMS prediction error also passes while compounding harm, which would be the strongest argument that dynamic, disparity-based evaluation — not device-level accuracy — is the standard autonomous ICU control actually needs.