When a hospital or health system rations a scarce treatment across patients whose disease evolves continuously, it faces two questions at once. The first is how much it cares about equity versus total health gained. The second is what the treatment will actually do, given that the only knowledge comes from sparse, irregularly timed records. The fairness literature answers the first with a tunable dial. The best known is α-fairness, a family of objectives in which one number α slides the goal from "maximize total benefit" (α = 0) through "proportional fairness" (α = 1) to "help the worst-off group first" (α → ∞). Neural ordinary differential equations (ODEs), networks that learn the rate of change of a hidden patient state so it can be evaluated at any time, answer the second. This note argues that combining the two naively produces a specific and dangerous failure. As α rises, the objective puts its weight on the worst-off group, and that group is typically the one with the least data. The optimizer therefore gets steered toward exactly the region where the learned dynamics model is least trustworthy. The resulting policy can report large gains for the disadvantaged group that exist only in the model.
We propose to measure this effect and to fix it. The fix is a conservatism penalty, a cost on policies whose predicted patient trajectories stray from the observed data. The penalty is computed separately for each group and scaled by that group's current weight in the welfare objective, so caution rises automatically wherever the fairness dial concentrates attention. A single policy network takes α as an input and is trained across the whole range of the dial. The deliverable is a fairness–efficiency trade-off curve whose fair end can be trusted. Today such curves are drawn using the model's own predictions, which overstate the benefit to minority groups.
Core Claim
Let $y_g(\pi)$ be the mean positive health benefit that group $g$ receives under allocation policy $\pi$, for example tumor-volume reduction relative to no treatment. Let $U_\alpha$ be the isoelastic utility used in Group-Fair Online Allocation in Continuous Time, equal to $x^{1-\alpha}/(1-\alpha)$, or $\ln x$ at $\alpha=1$. The planner maximizes $W_\alpha(\pi)=\sum_g U_\alpha(y_g(\pi))$. The gradient of this objective weights each group by $U_\alpha'(y_g)=y_g^{-\alpha}$, which grows explosively for the worst-off group as α grows. Suppose $\hat y_g$ comes from a learned continuous-time model whose error for group $g$ scales with that group's data scarcity. Then the error in the welfare gradient is dominated by the scarcest disadvantaged group at large α.
Our first claim is empirical and falsifiable. Under uniform conservatism, the gap between the model-predicted and the true worst-group benefit grows monotonically with α. Our second claim is that the following objective closes most of that gap without giving up the efficient end of the curve:
$$ \max_{\pi}\; \hat W_\alpha(\pi) \;-\; \kappa \sum_g \tilde w_g(\alpha,\pi)\, D_g(\pi). $$
Here $\hat W_\alpha$ is welfare predicted by the learned model. $D_g(\pi)$ is a path-space discrepancy, the signature-kernel maximum mean discrepancy of Conservative Continuous-Time Treatment Optimization, between simulated group-$g$ trajectories under $\pi$ and observed group-$g$ trajectories. $\tilde w_g$ is the normalized marginal welfare weight $\hat y_g^{-\alpha}/\sum_h \hat y_h^{-\alpha}$, held fixed (stop-gradient) during each update. $\kappa$ is a single global strength. The penalty spends its caution where the objective spends its attention.
Relation to Prior Work
Three papers sit closest, and each lacks one ingredient. Fairness in Survival Analysis with Distributionally Robust Optimization already places a fairness dial inside a survival neural ODE (SODEN). Its dial, however, is a subpopulation-size threshold for worst-case risk, not a welfare curvature. It also predicts static time-to-event outcomes and makes no allocation decisions. Group-Fair Online Allocation in Continuous Time supplies the α-fair objective in continuous time. It assumes stationary renewal processes with no evolving latent state, and it proves the exact problem PSPACE-hard, which motivates learned approximate policies rather than exact ones. Conservative Continuous-Time Treatment Optimization supplies our conservatism machinery and our simulators. It shows that offline continuous-time controllers exploit data-sparse regions, but its objective is single-patient and purely utilitarian.
Our contribution is to show that fairness weighting and model exploitation interact, and to couple them. None of the three studies that interaction.
Other papers constrain the design rather than supply it. Fair Continuous Resource Allocation with Equality of Impact bounds unknown concave response curves with secants and Gaussian processes. Its approach is limited to two groups and static rounds, which is why we need a learned dynamics model. Continuous-time Model-based Reinforcement Learning gives ensemble neural ODEs and continuous-time actor-critic training. We reuse its ensembles as a rival form of pessimism. MIMIC-IF motivates the premise that models perform worse for higher-risk demographic groups in real intensive-care data. That is the condition under which the failure we describe should appear.
Research Program
The dynamics model is a controlled latent ODE in the style of Latent ODEs for Irregularly-Sampled Time Series. Its encoder runs over irregular observations to infer an initial hidden state, and treatment doses enter the learned vector field as control inputs. Group membership is an input to the model, but the model is trained on data in which minority groups are deliberately underrepresented.
The policy observes the patient's inferred hidden state, the group, the remaining shared budget, and α. It outputs a dose schedule. We train it by backpropagating through the ODE solver over a population of simulated patients drawn from the learned model, sampling α from a log-uniform range in [0, 20] at each step. The single shared budget couples patients, which is what makes this an allocation problem rather than a set of independent ones.
The analysis has three parts. First, we measure the predicted-versus-true worst-group gap as a function of α under no conservatism, under uniform conservatism, and under welfare-weighted conservatism. Second, we test whether that gap tracks the per-group sample size. Third, we check whether one α-conditioned network matches a grid of separately trained networks.
Baselines
The naive reference point is a utilitarian policy (α = 0) trained on the same latent ODE without conservatism, alongside the no-treatment and uniform-split allocations. These show how much any fairness dial moves outcomes at all.
The strongest existing method is the signature-regularized controller of Conservative Continuous-Time Treatment Optimization. We extend it to the shared-budget α-fair objective with a single uniform penalty. On the lung cancer simulator it reports a ground-truth cost of 73.42, against 280.97 for TE-CDE, and a Spearman correlation of 0.40, against 0.29, between predicted and true policy costs. We first reproduce the single-patient 73.42 setting to validate our implementation. Note, however, that no reported result was found on Lacuna for any group-fair or shared-budget variant of this simulator, so all fairness comparisons are new measurements. For a fair comparison, both methods use the same dynamics model, data, solver tolerances, and number of gradient steps. Each method's κ is tuned on the same budget of 20 configurations, selected on held-out simulator validation seeds. We count this baseline as beaten if, at α ≥ 5, the true worst-group benefit is higher at equal or better true total benefit, and if the predicted-versus-true worst-group gap shrinks by at least half.
A second competitor is pessimism through ensemble disagreement, using the ensemble neural ODEs of Continuous-time Model-based Reinforcement Learning and penalizing by the lower confidence bound of group benefit. It gets the same tuning budget.
A third is a static fair allocator in the spirit of Fair Continuous Resource Allocation with Equality of Impact. It divides the budget once among groups using predicted response curves, and each group then runs an independent utilitarian schedule.
The key ablation keeps everything fixed but sets $\tilde w_g$ to the uniform value, isolating whether welfare weighting of the penalty is what matters.
Why not just…?
Why not just turn up a uniform conservatism penalty? A large enough uniform κ will also suppress hallucinated minority-group gains. The cost is pinning the majority group's policy to the observed treatments, which gives up efficiency at every α. This is a real threat if the efficiency cost is small. The ablation settles it by comparing the two methods' true fairness–efficiency curves, each at its best κ. If uniform penalties trace the same curve, we will report that the coupling is unnecessary.
Why not just use a discrete-time recurrent model on binned data? The irregular-sampling results in Latent ODEs for Irregularly-Sampled Time Series report mean squared error of 2.361 against 3.215 on PhysioNet. Both simulators also mask 30% of observations. Still, the exploitation effect we study is not specific to ODEs. We will run the full study with a GRU dynamics model as well. If the GRU matches our results, the contribution is the welfare-weighted penalty, not the ODE, and we will say so.
Why not just allocate the budget statically across groups and treat each patient optimally? Static division ignores that when a patient is treated changes their later state and future need, so budget committed early cannot follow patients who deteriorate. The static baseline tests this directly. We expect it to be competitive when treatment timing barely matters and to fall behind on the lung cancer horizon, where dose timing drives tumor volume.
Why not just train a separate policy for each α? That is the grid baseline, and it is fine if the per-policy compute is acceptable. The α-conditioned network is a convenience. We claim only that it matches the grid, and we report the comparison either way.
Proposed Experiments
We build group-structured versions of the lung cancer PK/PD and COVID-19 simulators from Conservative Continuous-Time Treatment Optimization. Each has three to five patient groups whose simulator parameters differ: the high-need group responds more slowly to treatment and also has a higher untreated progression rate. We generate offline data in which the high-need group makes up 5%, 10%, or 30% of trajectories, logged under a behavior policy that undertreats it. This mirrors the ventilation disparity found in MIMIC-IF. A shared dose budget couples all patients.
The metrics are true total benefit, true worst-group benefit, the predicted-minus-true worst-group gap, and the area between the true and predicted fairness–efficiency curves. All are averaged over at least ten seeds with confidence intervals.
The decisive check is the gap-versus-α plot at 5% minority representation. Our hypothesis predicts a rising gap under uniform conservatism and a flat one under welfare-weighted conservatism, with the difference shrinking as representation rises to 30%. If the gap does not grow with α even without conservatism, the premise is wrong and the proposal fails cleanly.
As a real-data sanity check that cannot provide counterfactual ground truth, we fit the latent ODE to MIMIC-IV and report per-group predictive error against group size. This tests whether the error pattern the method relies on exists outside simulation.
References
- Group-Fair Online Allocation in Continuous Time
- Conservative Continuous-Time Treatment Optimization
- Fairness in Survival Analysis with Distributionally Robust Optimization
- Fair Continuous Resource Allocation with Equality of Impact
- Continuous-time Model-based Reinforcement Learning
- MIMIC-IF: Interpretability and Fairness Evaluation of Deep Learning Models on MIMIC-IV Dataset
- Latent Ordinary Differential Equations for Irregularly-Sampled Time Series