A single case report shows that something can happen to one person. It cannot say how often it happens, or whether it would happen to the next patient. Clinicians are increasingly asked to make this jump in reverse. A population-trained model makes a recommendation for an individual, and someone must decide whether to trust it or to pay for individual evidence. That evidence might be an N-of-1 crossover trial (one patient alternates between treatments over several periods so their own response can be measured) or an expert review of the model's call. Individual evidence is expensive, and the capacity to produce it is fixed: monitoring slots, coordinator hours, specialist attention. The current proposal for this decision, in Personalization of Large Foundation Models for Health Interventions, uses a fixed uncertainty cutoff ($\sigma > 0.25$). That rule ignores how much capacity remains and how many patients are still coming.
We propose to treat escalation as an online budgeted allocation problem and to solve it with shadow pricing. A shadow price is a running "cost per unit" of the scarce capacity, raised when capacity is being used too fast and lowered when it sits idle. A patient is escalated when the expected value of individual evidence for that patient exceeds the current price of the capacity the test would consume. Two choices make this more than a port of known allocation methods. First, the value side is the value of information: how much the individual test is expected to change the decision. It is not the predicted benefit of treatment. Second, admission is randomized through an entropy-regularized price, so every patient's escalation probability is known. That probability turns the pile of escalated cases, which is otherwise a biased, case-report-like record, into data the population model can learn from without selection bias.
Core Claim
Standard dual pricing, as formalized by Dual Mirror Descent for Online Allocation Problems, scores each arrival by reward minus the price-weighted resource it consumes. It updates the price by mirror descent on the gap between actual consumption and the target burn rate, and guarantees $O(\sqrt{T})$ regret against the hindsight optimum under i.i.d. arrivals. Applied naively to patients, "reward" becomes expected benefit. That reproduces the "rational abandonment" of The Shadow Price of Reasoning, which assigns zero budget to items whose expected payoff is too low. In a clinic, this sheds exactly the unusual, high-cost patients that case reports such as Tanaka et al. and Zhou et al. show the population average fails on. Our claim has three parts. Replacing benefit with value of information removes this failure. A soft, entropy-regularized acceptance rule keeps the pacing guarantee while producing known propensities. Together these beat both fixed cutoffs and benefit-priced dual policies once arrival rates and patient mix shift over time.
The decision rule for patient $t$ is
$$ P(\text{escalate}_t) = \sigma\!\left(\frac{V_t - \lambda_t c_t}{\tau}\right), \qquad \lambda_{t+1} = \max\!\left(0,\ \lambda_t + \eta\,(c_t\,\mathbb{1}[\text{escalate}_t] - \rho)\right). $$
Here $V_t$ is the expected reduction in decision loss if patient $t$'s individual outcome were observed, computed from the population model's posterior. $c_t$ is the capacity the test consumes. $\lambda_t$ is the shadow price. $\rho$ is the per-arrival budget rate, $\eta$ is a step size, and $\tau$ is a temperature that controls how random admission is; the outer $\sigma$ is the logistic function. The temperature borrows the smoothing idea of Multi-slots Online Matching with High Entropy, which uses entropy to damp price oscillation. We use it for a different purpose: to keep every propensity bounded away from zero.
Research Program
The method has three parts. A population model outputs a posterior over each patient's response or label, and from it a value-of-information score. A pricing layer paces escalations against capacity and handles two-sided constraints. Following Joint Online Learning and Decision-making via Dual Mirror Descent and Nearly Optimal Competitive Ratio for Online Allocation Problems with Two-sided Resource Constraints, lower bounds enter as negative prices that subsidize escalation for subgroups the record would otherwise miss. A learning layer refits the population model on escalated cases with inverse-propensity weights.
Theoretically, we will extend the mirror-descent regret argument to the temperature-smoothed rule. The expected result is an extra $O(\tau T)$ term from smoothing, which suggests choosing $\tau \propto T^{-1/2}$ to preserve the $O(\sqrt{T})$ rate. We will also bound the variance of the weighted population estimate by the smallest propensity.
Empirically, we test where the method should and should not help. Under stationary arrivals, the hindsight-best fixed price is essentially the fluid optimum, so adaptivity alone should buy little. We predict that the gains come from shifting arrivals and from the value-of-information score, and the experiments are built to separate those two sources.
Baselines
The naive reference is never escalating, which means acting on the population model alone. Its counterpart is escalating at random at the budget rate, which is fair because it spends identical capacity.
The fixed-cutoff rule of Personalization of Large Foundation Models for Health Interventions escalates whenever $\sigma > 0.25$. It is the only existing method for exactly this decision. We run it both at 0.25 and with the cutoff tuned in hindsight to exhaust the budget exactly. That hindsight-tuned version is the strongest simple competitor, and beating it under shift is the central test. No reported result was found on Lacuna for this rule on any benchmark; the file gives only an illustrative migraine simulation with a 90% posterior.
The strongest existing method is benefit-priced dual mirror descent. It is the Dual Mirror Descent policy with reward set to predicted benefit, which reached roughly 90% of offline optimal reward within 10,000 iterations on the display-advertising generative model. A relative of it, the optimistic dual policy SP-UCB–OLP of A Two-Layer Framework for Joint Online Configuration Selection and Admission Control, reached 97.4% of oracle against 84.8% for greedy on the Alibaba Cluster Trace. These figures come from other domains and are not directly comparable. They show that the dual machinery itself is well tuned.
All policies receive the same population model, the same patient streams, the same capacity, and the same 50-configuration hyperparameter search for their tunable parameters on held-out streams. We also run periodic batch re-solving of the escalation linear program each simulated week and the hindsight offline optimum as an upper bound.
Two ablations isolate the key ingredients. The first swaps value of information for predicted benefit inside our pricing layer. The second sets $\tau \to 0$, removing propensities. Beating a baseline means lower decision loss at equal budget use, with no worse budget violation, across at least 80% of shifted stream orderings.
Why not just…?
Why not just use a fixed uncertainty cutoff tuned to the budget? Under stationary arrivals this is nearly optimal, and we expect it to roughly tie our method there. The evidence that adaptivity matters comes from other domains. Static rules lost about half the reward in A Two-Layer Framework for Joint Online Configuration Selection and Admission Control, and round robin was 45.75% worse on average latency in Online Linear Programming for Multi-Objective Routing in LLM Serving. Neither result is clinical. Our arrival-shift experiments settle the question directly: if the tuned cutoff matches us under shift, the pricing layer is unnecessary and the contribution reduces to the value-of-information score.
Why not just apply standard dual pricing with predicted benefit? It paces capacity correctly but spends it on patients whose decision would not change, and it abandons high-variance outliers. The first ablation measures this directly. We also report outcomes separately for the heavy-tailed outlier subgroup.
Why not just add an optimism bonus, as UCB-style methods do? Optimism about reward explores arms, not individuals. Inside the escalation decision it inflates the benefit estimate rather than measuring whether a test would change the action. We include benefit-plus-UCB as an additional row.
Why not just escalate a random sample? Random escalation gives unbiased calibration but wastes capacity. Our method contains this as the $\tau \to \infty$ limit, and we report the frontier between decision loss and calibration error as $\tau$ varies.
Proposed Experiments
The first testbed is a simulated N-of-1 clinic. Patient treatment effects come from a mixture with a 2–10% heavy-tailed outlier component, deliberately violating the Gaussian smoothness that lets greedy dual policies explore "for free" in Smoothed Adversarial Linear Contextual Bandits with Knapsacks. Each crossover trial occupies monitoring slots for several weeks. Arrival rates and the outlier fraction drift over a two-year horizon. The metrics are regret against the hindsight optimum, budget overrun, outlier-subgroup loss, and calibration (expected calibration error) of the refit population model on non-escalated patients.
The second testbed uses real labels from the n2c2 2018 cohort-selection data: 288 patients and 13 criteria, with expert agreement of Cohen's $\kappa = 0.54$. The population model is zero-shot LLM screening, following Zero-Shot Clinical Trial Patient Matching with LLMs. That system reported +6 Macro-F1 over the prior best at an estimated $1.55 per patient, against $34.75 for manual screening. Escalation reveals the expert label at manual-review cost under a weekly reviewer budget. Lower bounds require minimum review coverage for each criterion-defined subgroup. We replay thousands of patient orderings, including orderings that front-load or back-load particular subgroups, and measure Macro-F1 at fixed review spend.
The decisive checks are three. First, under shifted orderings our method must beat the hindsight-tuned cutoff on decision loss without exceeding budget. Second, the value-of-information ablation must lose specifically on the outlier subgroup. Third, the inverse-propensity-weighted population model must be better calibrated than one refit on the escalated cases without weights. If the second check fails, the case-report critique does not translate into an allocation advantage, and we will report that.
References
- Personalization of Large Foundation Models for Health Interventions
- Dual Mirror Descent for Online Allocation Problems
- The Shadow Price of Reasoning: Economic Perspective on Optimal Budget Allocation for LLMs
- A RESECTED CASE OF ADVANCED ESOPHAGEAL CARCINOMA SHOWING A PROMINENT RESPONSE TO NEOADJUVANT CHEMOTHERAPY WITH 5-FU AND LEUCOVORIN
- Laparoscopic pancreaticoduodenectomy in A-92-older Chinese patient for cancer of head of the pancreas
- Multi-slots Online Matching with High Entropy
- Joint Online Learning and Decision-making via Dual Mirror Descent
- Nearly Optimal Competitive Ratio for Online Allocation Problems with Two-sided Resource Constraints and Finite Requests
- A Two-Layer Framework for Joint Online Configuration Selection and Admission Control
- Online Linear Programming for Multi-Objective Routing in LLM Serving
- Smoothed Adversarial Linear Contextual Bandits with Knapsacks
- Zero-Shot Clinical Trial Patient Matching with LLMs