acceptodds
Under review as a conference paper at ICLR 2027

Beyond Patient-Level Triage: Pre-Arrival Care Navigation under Shared, Limited Capacity

Abstract

Pre-arrival care navigation, recommending a care setting from a patient's pre-arrival description, offers an opportunity to direct patients to appropriate care before emergency-department (ED) resources are committed. EDs often care for patients whose needs could be met in lower-acuity settings, yet emergency triage largely assesses acuity and resource needs after ED arrival, when care-routing options have already narrowed. The decision serves two interests: each patient seeks capable and timely care, while the health system must keep care timely for patients with urgent needs; under constrained capacity, lower-acuity demand competes with later, more urgent arrivals. The patient's interest can be scored at every routing decision, but the system's interest is determined by a whole run of decisions and observed only after them, so it is hard to learn from single decisions. We introduce Stackelberg-Inspired Commitment Learning (SICL), a hierarchical framework for language-model routing policies that gives the system's interest a decision of its own. A system-level routing posture, held across a segment of arrivals, sets site-level incentives; each patient's destination is then chosen in the patient's own interest under those incentives; and during training the posture is credited with the segment's system-level outcome, according to how closely the routing followed it. We further construct TriageRoute, a dynamic emergency-care benchmark derived from MIMIC-IV-ED v2.2 that models heterogeneous care sites, travel delays, evolving queues, and workload-sensitive outcomes, where every policy sees the same patient information and care-network state and is given the same instructions. Despite identical information, policies differ widely in timely access for patients with urgent needs; in controlled simulation, SICL achieves the highest timely access among policies given the same information while also achieving the highest patient utility, at the cost of reduced capacity efficiency.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language model agents, sequential decision making, hierarchical reinforcement learning, credit assignment, simulation environment, clinical decision support

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vj8ZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies pre-arrival care navigation under shared, limited capacity, where routing decisions must account for both individual patients’ needs and their effects on subsequent arrivals. The authors introduce TRIAGEROUTE and the SICL. Experiments show improved urgent, timely access compared with the evaluated LLM-based policies, with a trade-off in capacity efficiency.

Overall, I think the direction is very promising, but the main concerns are insufficient benchmark analysis and the need to strengthen the related work comparison. If the author could resolve those, I would increase the score.

Critical strengths
  1. Individually reasonable recommendations can collectively overload scarce medical resources and delay urgent patients. Moving beyond isolated patient-level evaluation toward sequential decisions under shared capacity is practically meaningful.
  2. The benchmark is interesting and potentially useful. Combining clinical descriptions with heterogeneous care sites, travel times, and evolving queues enables evaluation of consequences that static triage benchmarks cannot capture.
Critical weaknesses
  1. TRIAGEROUTE is a central contribution, but much of the information needed to assess its construction and validity is deferred to the appendix. The main paper should devote more space to case selection, clinical-label construction, the care-network configuration, and validation of the simulator.
  • In this one, several consequential choices remain insufficiently justified, including the six-site network and the treatment of recorded resource use as required care. Please see the details in the question part
  1. At an abstract level, the task resembles traditional online allocation or routing with heterogeneous resources, uncertain demand, capacity constraints, and congestion externalities. What are the main differences between the health-care area and those domains? The paper should clarify which characteristics of the medical setting create technical challenges beyond this established formulation.

  2. The paper includes LLM baselines with the same observations, but the classical routing references receive ground-truth acuity and resource needs, whereas SICL must infer them. A key missing baseline is a modular pipeline in which an LLM receives the same inputs as SICL, predicts acuity and resource requirements, and passes these predictions to a conventional routing algorithm. This comparison would help determine whether SICL’s gains require its proposed learning mechanism.

Questions for authors
  • Care-network configuration and resource labels. Why does the benchmark use exactly six care sites? Could the authors evaluate additional configurations? Furthermore, what validates the extracted resource requirements?

  • Of the 69,043 ED stays linked to discharge summaries, 6,166 become detailed record-grounded cases. How many records were retained or excluded at each stage, and for what reasons?

  • The 1,800 supplementary low-acuity cases follow a different construction process and have a thinner information schema. Could these differences introduce textual shortcuts that allow a model to identify low-acuity cases from presentation style or information completeness rather than clinical content? Has this possibility been tested?

  • Clinical validation of generated content. Were the reconstructed interviews and extracted or inferred resource requirements reviewed by clinicians on a representative sample? If so, what were the error rates and agreement levels?

  • The motivating example emphasizes preserving high-capability resources by diverting lower-acuity patients. However, SICL routes 75% of ESI 3–5 patients to high-capability sites, compared with 69% for PPO. Which specific routing behaviors improve urgent, timely access rate?

6 comments
AuthorsOpenReview ↗

Response to Reviewer vj8Z (1/4)

Thank you for the careful review. You named two concerns as decisive: the analysis of the benchmark and the comparison with related work. We will add a main-text section on how TriageRoute is built and validated (case funnel, label construction, clinician review, care network), a related-work paragraph on online allocation, and a comparison with existing clinical LLM benchmarks. We also ran the modular baseline you suggested. We answer W1 and Q1–Q3 first, then related work (W2) and the modular baseline (W3), and finally the clinician review (Q4) and Q5.

W1. Benchmark construction in the main text.

We agree this material belongs in the main text. We will move it into a new section, using the additional page, that covers case selection (the funnel in Q2), label construction (Q1), the clinician review (Q4), the care network (Q1) and the simulator's checks (the arrival profile against NHAMCS, and the capacity variants in Q1). Two choices you flag need a direct justification.

  • Recorded resource use as required care. The resource label is the set of tests and procedures documented in the ED course, matched to a catalogue of 29 early-ED resources defined with collaborating physicians. It records the care the patient actually received, so it is reproducible across thousands of cases instead of depending on one reviewer's judgment. The ESI algorithm itself separates levels 3–5 by the number of resources a patient is expected to need (Gilboy et al., 2023), and we remove cases whose documented resources contradict the nurse-assigned level (Q2). In the simulator the label is used only for feasibility, that is, whether a site can deliver the documented care.
  • Six sites. See Q1.

Q1. Six sites, other configurations, resource validation.

  • Why six. The care-network layer is configurable (App. C.9): a configuration lists the sites of a region with tier, capability profile, effective capacity and hours. The six-site network is one instance, anonymized from public information on a real metropolitan region. It is the smallest network that contains every tier of a regional emergency-care system: two tertiary centers, so that there is a choice between equally capable sites, a general and a community hospital, an urgent-care center, and a clinic.
  • Other configurations. The tension we study is set by how scarce capacity is, so we evaluated variants that scale effective capacity to ×0.8 and ×1.2 with the sites fixed (Table R2 in the general response). UTA moves with scarcity for every policy. Among policies that see SICL's inputs, SICL is highest at all three levels (0.446, 0.540, 0.557), and its margin over PPO narrows when beds are ample (+0.037 at ×1.2, against +0.061 at ×0.8 and +0.101 at the default level). Configurations for other regions use the same schema.
  • What validates the resource requirements. Four checks. (i) The catalogue of 29 early-ED resources was defined with collaborating physicians. (ii) Each extracted resource carries a clinical rationale, and 80% of them quote the note; checked against the source notes, 90% of these quotes appear verbatim (up to whitespace and punctuation). A test whose ED timing cannot be confirmed is left out of the label. (iii) Cases whose documented resources contradict the nurse-assigned ESI are removed (Q2). (iv) Two physicians independently reviewed the pre-arrival presentations and acuity (Q4).
AuthorsOpenReview ↗

Response to Reviewer vj8Z (2/4)

Q2. Case funnel from 69,043 to 6,166.

The counts below come from the construction scripts; we will show them as a funnel in §5.1. The split by patient comes before case generation: no evaluation encounter appears in the pool, and record-grounded pool cases share no patient with the evaluation encounters.

StageED staysPatientsKept when
MIMIC-IV-ED v2.2425,087205,504–
Linked to a discharge summary69,04352,344the stay's admission has one ED stay and one MIMIC-IV-Note discharge summary
Repeat-visit patients54,4828,001the patient has more than three ED stays, so earlier encounters can be shown
ED-course extraction42,2206,344extraction of the ED course succeeds and acuity is recorded
Acuity–resource consistency36,5066,334documented resources do not contradict the nurse-assigned ESI
Split by patient7,187 pool / 29,319 held out1,266 / 5,068no patient on both sides
Record-grounded cases6,1661,255interviews generated and pass validation

Discharge summaries longer than 10,000 tokens are also excluded along the way. The 1,800 supplemented low-acuity cases (App. C.4) are added on the pool side, giving 7,966 cases from 2,650 patients.

Q3. Text shortcuts in the supplemented cases.

We tested this in two places, and the answer differs between the pool and the evaluation.

  • No evaluation arrival is a supplemented case. Evaluation streams are drawn only from held-out record-grounded encounters (Q2); supplemented cases exist only on the pool side. Every evaluated low-acuity arrival therefore has the full record-grounded schema, and its source cannot be read from its format.
  • Same input template for every case. Every policy reads every case through the same template, with the same fields in the same order. No policy sees source text, construction prompts or the source of a case.
  • Tested on the evaluation inputs. A logistic-regression probe that uses only missingness, structural counts and text lengths, with five-fold cross-validation grouped by patient, identifies ESI-4 cases at balanced accuracy 0.53 (label-permuted controls 0.43–0.57) and AUROC 0.57.

In the pool, where both sources coexist, the two formats are distinguishable. The same probe trained on the pool and applied to the evaluation stream does not detect ESI-4 arrivals (AUROC 0.55), so the format signal does not carry over to record-grounded encounters. We will state the schema difference and these checks in §5.1, and release a source flag with every pool case so that users can stratify or reweight by source.

AuthorsOpenReview ↗

Response to Reviewer vj8Z (3/4)

W2. Difference from online allocation.

We agree that, abstractly, TriageRoute is an online allocation problem with heterogeneous resources and congestion externalities. When the type is revealed, classical methods are strong: given recorded acuity and resource needs, Dual pricing (bounded) reaches 0.513 UTA (Table 7). Three characteristics of the medical setting go beyond that formulation.

  • The type is not revealed. Online allocation assumes that the arriving item's type, or a calibrated signal of it, is observed. Here acuity and resource needs are observed only through a noisy, language-based pre-arrival interview. The modular baseline (W3) measures the cost: the same routing rules that reach 0.50–0.51 with recorded acuity reach 0.44–0.47 when an LLM infers acuity and resources from the inputs SICL sees, while SICL reaches 0.540 on the same inputs.
  • The routed patient's interest is an objective, not a cost. A dispatcher maximizes a system objective; a navigator also owes each patient a capable, close and timely site. We therefore score patient utility beside system outcomes, and destinations are chosen in the patient's interest (§4).
  • The system objective is thresholded, delayed and set by runs of decisions. Whether an urgent patient reaches a bed within 30 minutes depends on placements made earlier for other patients. Dual pricing handles this well when types are revealed (Table 7). When the router is a language policy reading interviews, the open problem is to credit a delayed, shared outcome to the decisions that shaped it, so we see the learning problem mainly as one of credit assignment rather than dispatch.

We will add a related-work paragraph on online matching and resource allocation, network revenue management, and patient scheduling and ambulance dispatch, organized around these three differences.

W3. Same-input modular baseline.

We agree this is an important comparison and have run it (Table R1 in the general response). An LLM receives exactly the inputs SICL receives, predicts ESI and the resource set, and passes them to the two strongest classical routing rules of Table 7, Shortest-delay and Dual pricing (bounded), each in its own simulator run. We use two LLMs: a prompted Qwen3-32B, and a Qwen2.5-1.5B (SICL's backbone) fine-tuned on our training labels. The 32B model's prompt was chosen among three versions on 150 training-side cases by a rule fixed in advance, and frozen before any evaluation run; the fine-tuned model uses its own training prompt.

  • With recorded acuity and resources, the two rules reach 0.504 and 0.513 UTA (Table 7). With the clinical state inferred by Qwen3-32B, they reach 0.441 ± 0.029 and 0.456 ± 0.065. SICL reaches 0.540 ± 0.030 on the same interviews (our reply to Reviewer MxN4, W4), with higher patient utility (0.679 vs 0.656 and 0.666) and lower capacity efficiency, the trade-off reported in §6.2.
  • The 32B model finds most urgent patients (urgent recall 0.71) but over-triages: it labels 41% of arrivals ESI 1–2, against 13% in the records, and rarely assigns ESI 4–5. The rules then hold capacity for patients who do not need it, and incomplete resource predictions raise the transfer rate to 4–6%, against 1% for SICL.
  • The fine-tuned 1.5B model finds more urgent patients (urgent recall 0.86) but over-triages further, labeling 54% of arrivals ESI 1–2. With it the rules reach 0.436 ± 0.049 and 0.470 ± 0.010 UTA, with patient utility 0.651 and 0.661, again below their results with recorded acuity and below SICL.

These results suggest that SICL's gain is not reproduced by pairing a prompted or fine-tuned LLM with a strong routing rule, and that learning the routing from system outcomes matters.

Reviewer_X1UeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to route patients to care sites before arrival, while accounting for the capacity available to later patients. It proposes SICL, where a leader selects a routing posture and a follower chooses a destination for each patient. Response-Attributed Credit (RAC) adjusts the leader’s training signal based on the system outcome and how closely the routing follows the posture. The authors also introduce TriageRoute, a simulation benchmark built from MIMIC-IV-ED records. SICL improves urgent timely access and the specified patient utility over PPO, with a reduction in capacity efficiency.

Critical strengths

The problem is relevant. Looking at patients independently can miss how one routing decision affects the care available to others. Studying this interaction is a useful direction beyond predicting triage labels.

The results within the tested setting are promising. SICL improves urgent timely access from 0.458 to 0.532 over PPO. The same-information protocol is useful, and the comparisons include both prompted and trained policies

Critical weaknesses

Methodological contribution. The method looks fairly incremental on top of HiPER. Splitting system-level and patient-level objectives makes sense, but most of the hierarchical structure is already in place, so the actual additions feel limited. The proposed RAC could be useful and new, but the current ablations don't show whether the improvement comes from response-based attribution itself, or just from the extra segment-level feedback.

  • The evaluation needs stronger validation. The evaluation lean heavily on the authors' own benchmark with mostly custom metrics. Using MIMIC data is fine, but it doesn't really validate the reconstructed inputs, the simulator, or the scoring choices you've made. The threshold and temperature experiments are helpful, but they stop short of discussing what happens when patient demand or hospital capacity shifts. Appendix C.8 mentions such scenarios, yet the actual evaluation is left to future work.

  • Some metric choices make the claimed benefits hard to interpret. UTA excludes the initial trip and only measures waiting after arrival, so better UTA does not necessarily mean that patients receive care sooner overall. The patient utility also gives a strong reward for choosing higher-tier hospitals.

  • The evidence across models is limited. SICL is trained only on Qwen2-1.5B-Instruct and it is not clear on whether the method is effective on other LLMs. Some important implementation details are also missing, including the actual posture priors, several RAC parameters, and a complete description of the simulator’s service dynamics.

Questions for authors
  1. What if we use an LLM for clinical assessment and then use a classical dispatcher to fulfill the task?

  2. Regarding the patient utility, can the authors justify the strong preference for higher-tier hospitals and show whether the reported patient–system conflict and SICL’s gains persist under weaker tier preferences or stronger waiting penalties?

4 comments
AuthorsOpenReview ↗

Response to Reviewer X1Ue (1/2)

Thank you for the detailed review. We answer each point below.

W1. Increment over HiPER; attribution vs. extra feedback.

SICL keeps HiPER's training machinery and changes what each level is for:

HiPERSICL
Upper levelfree-text subgoalsystem-level routing posture that commits site-level incentives
Lower level optimizestask rewardthe routed patient's payoff under those incentives
Upper level credited withepisode returnthe segment's system-level outcome, weighted by how closely routing followed the posture (RAC)

On the same backbone and inputs, HiPER reaches 0.246 UTA and SICL 0.532 (Table 2). Table 3 shows what the changes contribute when added in turn: with the posture and its incentives but without RAC, UTA stays close to PPO's level, and it rises above PPO only once RAC is added.

On attribution versus extra feedback, you are right that removing RAC removes both the segment-level outcome signal and the response weighting, so the current ablation cannot separate the two. We are running the control that does: it keeps the segment-level signal and fixes the response weight to one. For the final version we will add (1) this control and (2) results over several training seeds for each ablation row.

W2. Benchmark validation; demand and capacity shifts.

We added evidence on each part.

  • Inputs and labels. We report the case funnel stage by stage, from 69,043 MIMIC-IV-ED stays linked to discharge summaries to 6,166 record-grounded cases. The split by patient comes before case generation: no evaluation encounter appears in the case pool, and record-grounded pool cases share no patient with the evaluation encounters. Resource labels are the workup documented in the ED course; 80% of extracted resources quote the note, and 90% of these quotes appear verbatim in the source notes. Two physicians with more than ten years of practice independently reviewed record-grounded cases: their acuity matches the nurse-assigned level exactly in 58% of cases, on par with the 59% at which ESI-trained nurses assign the reference level in standardized scenarios, and within one level in 84–100%, and 92–96% of interview questions were rated essential or optional. Full details are in our reply to Reviewer vj8Z (Q1–Q4).
  • Capacity shifts. Following your suggestion, we evaluated the trained policies at ×0.8 and ×1.2 bed capacity, keeping sites, capabilities and arrivals fixed (Table R2 in the general response). Among policies that see SICL's inputs, SICL gives the highest UTA at all three levels: 0.446, 0.540 and 0.557, against 0.385, 0.439 and 0.520 for PPO and at most 0.381, 0.456 and 0.528 for the Qwen3-32B pipelines. Its margin narrows when beds are ample and capacity is less contested.
  • Demand. Load depends on demand relative to capacity, so the ×0.8 setting also corresponds to 25% more arrivals per bed. Across ten additional held-out days with different arrival streams, SICL exceeds PPO on all ten (our reply to Reviewer MxN4, W4).

W3. UTA and the initial trip.

Thank you for catching this, and we apologize for the confusion. The description in §3 and App. A.1 is inaccurate. In the code, the wait that enters UTA runs from the routing decision, when the patient enters the system, to bed assignment, so it includes the initial trip. We will correct both passages. This is also the clock that matters for pre-arrival navigation, because the routing decision sets both the trip and the queue; care for time-critical emergencies is likewise timed from first medical contact, transport included. Better UTA therefore does mean that urgent patients reach a treatable bed sooner overall: 33.4 vs 38.4 minutes on average under SICL and PPO (five runs each on identical interviews). The threshold sensitivity in App. D.4 uses the same decision-to-bed time. We answer the utility question under Q2.

AuthorsOpenReview ↗

Response to Reviewer X1Ue (2/2)

W4. One backbone; missing details.

We agree that a second backbone would strengthen the claim. Retraining the full pipeline did not fit in the discussion period, and we will add results with a second backbone in the final version. Scale alone does not close the gap: with the same information, prompted models from 4B to 120B parameters all stay below the trained 1.5B SICL policy (Table 2). The trained policies use Qwen2.5-1.5B-Instruct; we will correct the version name in the paper. We will add the missing details to the appendix. In brief:

  • Posture priors. Each of the four postures is a fixed reference distribution over sites, set by tier. Urgent-first concentrates mass on the tertiary centers; preserve-tertiary and spread-nonurgent move mass away from the two tertiary centers and spread it evenly over the other four sites; balance-load is close to uniform. The priors enter only the training-time incentive. At inference the posture is only text in the follower's context, and the follower may route to any site. The exact values will be listed in the appendix.
  • RAC constants. c1, c2, c3 = 1, 0.5, 0.3; EMA rate 0.02; λ target 0.3 with a 200-update warm-up; τ annealed from 2 to 1; g_min = 0.2; w ∈ [0.5, 1.5]; incentive β = 0.1.
  • Service dynamics. Each site has a bed capacity, reduced overnight, and a set of available resources. An admitted patient holds a bed until the documented ED workup is complete. Bedside items such as monitoring run in parallel, focused assessments run in sequence, and imaging and other serialized diagnostics are shared queues at each site, so a patient's time in bed depends on both their documented needs and the site's current diagnostic load. Waiting patients are served by ESI priority, first-come first-served within a level, and beds are not pre-empted. A patient whose documented needs exceed the site's capabilities is transferred on admission to the nearest capable site, and the transfer drive time is added.

Q1. LLM assessment plus a classical dispatcher.

We ran exactly this pipeline (Table R1 in the general response; details in our reply to Reviewer vj8Z, W3). With a prompted Qwen3-32B or a fine-tuned Qwen2.5-1.5B estimating acuity and resources from SICL's inputs, the two strongest classical routing rules reach 0.44–0.47 UTA, below their 0.50–0.51 with recorded acuity and below SICL's 0.540 on the same interviews. Both models over-triage, labeling 41% and 54% of arrivals urgent against 13% in the records.

Q2. Weaker tier preference, stronger waiting penalty.

The default profile models patients who prefer more capable hospitals; patients often bypass the nearest facility for a higher-level one (Radcliff et al., 2003; Liu et al., 2018). It is a configurable assumption, and we agree that the default profile weights the tier strongly (App. A.2). UTA itself is computed without the utility, but the trained routing depends on it, because the follower is rewarded with the patient's payoff. The direct test of your question is therefore to retrain under a different profile. We are retraining SICL and PPO under a time-sensitive profile, with the tier preference halved and the waiting penalty saturating at 4 hours, and will report UTA, patient utility and the share of ESI 3–5 patients sent to high-tier sites in the revision. Re-scoring the original trajectories under alternative profiles already leaves SICL's utility at or above PPO's in every case:

ProfileSICLPPO
Default0.6800.669
Tier preference halved0.6300.622
No tier preference0.5510.547
Weak tier preference, stronger waiting penalty0.4430.421

Values are comparable within a row, not across rows, since each profile is normalized to its own range.

Thank you again for the constructive review. We hope the capacity results, the clarification of what UTA measures and the same-input modular baseline address your main concerns. If they do, we would be grateful if you would take them into account in your final assessment. We are glad to answer any further question before the discussion closes.

Thank you for the detailed response and additional experiments. The new baseline and capacity experiments address some of my concerns.

You mentioned that the control with the response weight fixed to one is currently running. Also, you mentioned retraining SICL and PPO under a more time-sensitive utility profile. Is that possible to share the results before the discussion deadline? Thanks.

AuthorsOpenReview ↗

Response to Reviewer X1Ue

Thank you for the follow-up. We have completed the control experiment with the response weight fixed to one (with checkpoint selection on validation as for SICL, evaluated under the protocol of §6.3 across five decodes). Without response-based weighting, the leader collapses to keeping a single posture throughout the day and reaches 0.406 ± 0.050 UTA, compared to 0.540 ± 0.030 for SICL under the same evaluation protocol. This confirms that response-based credit attribution, rather than un-attributed segment-level feedback alone, is what enables the leader to dynamically adapt postures and improve urgent timely access. We will include this control together with the time-sensitive utility retraining in the revised version.

Reviewer_MxN4OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper provides a mechanism and simulator for triaging at care facilities using a stackelberg mechanism. It is an interesting problem. The paper is presented well.

Critical strengths
  1. Well motivated problem
  2. Experimental section is good and there is good diagnostic analysis
Critical weaknesses
  1. The motivation for Stackelberg framing is unclear. What is the sequential aspect of it? Why is there a need to use a game theoretic model?
  2. The source of the gain due to the new method is not clear.
  3. From my understanding, the objective was to preserve capacity for urgent patients. However, the SICL seems to send more ESI 3-5 patients to high capability sites than PPO (75% vs 69%) and its appropriateness (0.247) is among the worst. What am I missing here?
  4. For Table 8, across multiple runs do patients always arrive in the same order? What is the source of variance here? Overall, statistical significance of the results is unclear. 5.How many arrivals and urgent patients are in each run? Can you report results over multiple independent arrival streams and several training seeds per method?
  5. The performance of LLM acuity and classical router pipeline is unclear.
Questions for authors

Please see weaknesses above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer MxN4 (1/2)

Thank you for the careful review and the specific questions. We answer them in order.

W1. Why Stackelberg, and what is sequential.

Why a leader–follower (game-theoretic) model. Each routing decision serves two parties with different objectives. The system needs timely access for every urgent patient, while the routed patient wants a capable, close and quick site. The system cannot dictate a destination without overriding the patient's interest; it can only shape the choice. That is a leader–follower situation, and modeling it as one gives the system's objective a decision of its own:

  • the system-level leader first commits to a routing posture, which sets site-level incentives;
  • the follower then picks the destination in the routed patient's own interest under those incentives;
  • in training, the leader is credited through the response it induced: the segment's system-level outcome, weighted by how closely routing followed the posture.

A flat policy has no such decision, so the outcome of a whole segment has to be spread over individual destination choices, each of which also serves a patient. We use the Stackelberg structure as a design lens and do not compute an equilibrium, which is why we call the method Stackelberg-inspired.

What is sequential. The move order: the leader commits, then the follower responds. This holds within a decision (the posture is fixed before the destination is chosen) and across arrivals (a posture stays in force until the leader switches it, so one commitment shapes a run of routing decisions).

W2. Source of the gain.

Four comparisons, all with the same inputs at inference, narrow it down.

  • Inputs. An LLM that reads exactly SICL's inputs and hands its acuity and resource estimates to the two strongest classical routing rules reaches 0.44–0.47 UTA, below SICL's 0.540, whether it is prompted or fine-tuned (Table R1 in the general response; W5).
  • Reward. PPO is trained on the same episode-level urgent-access reward (§5.3) but spreads it over every decision of the day. Adding the posture and incentives does not change this by itself; UTA rises above PPO only once the segment's outcome is credited to the posture that shaped it (Table 3).
  • Where the gain appears. It is concentrated where capacity binds: when an urgent patient meets a saturated network (Table 8; W3), and when beds are scarce rather than ample. SICL's lead over PPO is +0.061 at ×0.8 capacity and +0.101 at the default level, and narrows to +0.037 at ×1.2 (Table R2).
  • Beyond one day. SICL exceeds PPO on all ten additional held-out days (W4).

These comparisons suggest that the gain comes mainly from where the system outcome is credited, not from extra inputs or from the reward itself.

W3. 75% vs 69%, and low appropriateness.

We will revise the motivating example so that it states the objective more precisely. The objective is timely access for urgent patients, and diversion is only one way to get there. Each destination is chosen in the routed patient's own interest under the posture's incentives, and the default preference profile values higher-tier sites. That is why SICL sends more ESI 3–5 patients to them, has the highest patient utility, and scores lower on appropriateness, which penalizes exactly those placements.

The gain for urgent patients comes from saturated moments. An urgent patient who meets a saturated network gets a bed within 30 minutes in 52.4% of cases under SICL and 43.8% under PPO; at other moments both reach 56.7% (Table 8). The routing behind this shows in where urgent patients go. In a re-evaluation with five runs on identical interviews (W4):

  • SICL sends 55% of urgent patients to the nearest capable site, against 26% for PPO, which cuts their mean drive from 13.2 to 9.7 minutes. Waits after arrival are also slightly shorter (23.7 vs 25.1 minutes), so urgent patients reach a bed in 33.4 vs 38.4 minutes on average.
  • PPO sends 26% of urgent patients to the general hospital, where only 35% reach a bed within 30 minutes. SICL sends 13% there and more to the second tertiary center.
  • Among lower-acuity patients, the 75% vs 69% gap comes entirely from ESI 3 (74% vs 57% to tertiary sites). For ESI 4–5, SICL sends fewer patients to tertiary sites than PPO does (77% vs 82%).

SICL therefore improves urgent access mainly through where urgent patients are placed, not by turning lower-acuity patients away from the sites they prefer. The cost is lower capacity efficiency, the trade-off reported in §6.2 and Appendix D.3.

AuthorsOpenReview ↗

Response to Reviewer MxN4 (2/2)

W4. Variance, counts, arrival streams and seeds.

We will state these details explicitly in the revised version. Within an evaluation day, patients and their arrival order are the same in every run, and the policies decode stochastically (temperature 0.1). Each run simulates a full day of the care network: 878 arrivals, 117 of them urgent (ESI 1–2), drawn from held-out encounters.

  • Decoding variance. We re-ran both policies five times on identical interviews. SICL reaches 0.540 ± 0.030 UTA and PPO 0.439 ± 0.024 (Welch's t-test, p < 0.001), in line with the reported 0.532 and 0.458.
  • Independent arrival streams. We evaluated ten additional held-out days, each with 829 to 919 arrivals and 110 to 122 urgent patients. SICL exceeds PPO on all ten, by 0.115 UTA on average (paired 95% CI 0.074 to 0.155; three runs per method and day), which is about 13 more urgent patients per day reaching a bed within 30 minutes. Patient utility is also higher on nine of the ten days.
  • Training seeds. We will report results over several training seeds per method in the final version.

W5. LLM acuity plus a classical router.

We ran this pipeline. The full description is in our reply to Reviewer vj8Z (W3), and the numbers are in Table R1 of the general response. When an LLM estimates acuity and resources from SICL's inputs, both routing rules fall below their results with recorded acuity and below SICL, whether the LLM is prompted or fine-tuned. The main cause is over-triage: the prompted 32B model labels 41% of arrivals urgent and the fine-tuned 1.5B model 54%, against 13% in the records, so the rules hold capacity for patients who do not need it.

We hope these answers clarify where SICL's gain comes from and how robust it is. If they address your concerns, we would be grateful if you would take them into account in your final assessment. We are glad to answer any further question before the discussion closes.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers and the AC

We thank the reviewers for their careful and constructive reviews. They found the problem well motivated and practically relevant (MxN4, X1Ue, vj8Z), the benchmark interesting and potentially useful (vj8Z), the same-information protocol useful (X1Ue), and the experiments and diagnostic analysis good (MxN4). Below we summarize the four main requests and our response to each.

Request 1: a modular baseline in which an LLM estimates acuity and resources and a classical routing rule chooses the site (all three reviewers).

Response. We ran it with two LLMs, a prompted Qwen3-32B and a Qwen2.5-1.5B fine-tuned on our training labels, each feeding the two strongest classical routing rules of Table 7 (Table R1; full details in our reply to Reviewer vj8Z, W3). With either LLM the rules reach 0.44–0.47 UTA, below their 0.50–0.51 with recorded acuity and below SICL's 0.540 on the same interviews. Both LLMs over-triage heavily.

Request 2: clearer evidence on how TriageRoute is built and validated (vj8Z, X1Ue).

Response. We report the full case funnel from MIMIC-IV-ED, the clinician review with agreement levels, a check of the extracted resources against the source notes, and tests for presentation shortcuts in the supplemented cases (our reply to Reviewer vj8Z, Q1–Q4). We will move this material into a new main-text section.

Request 3: a clearer account of where SICL's gain comes from (MxN4, X1Ue).

Response. The gain is not reproduced by pairing an LLM with a strong routing rule (Request 1), it appears only once the posture is credited through RAC (Table 3), and it is concentrated where capacity binds (Table 8 and Table R2). This is the pattern we would expect if the gain comes from crediting the delayed system outcome to the routing decisions that shaped it, rather than from better acuity inference (our replies to Reviewer MxN4, W2, and Reviewer X1Ue, W1).

Request 4: robustness beyond a single evaluation setting (MxN4, X1Ue).

Response. We scaled bed capacity to ×0.8 and ×1.2. SICL has the highest UTA among policies with SICL's inputs at all three levels, and its lead narrows when beds are ample (Table R2). On ten additional held-out days, SICL exceeds PPO on all ten (UTA +0.115 on average, paired 95% CI 0.074 to 0.155; our reply to Reviewer MxN4, W4).

Table R1. UTA of the two strongest classical routing rules of Table 7 when acuity and resource needs come from each source (five runs each). SICL reaches 0.540 ± 0.030 on the same interviews.

Acuity and resources fromLabeled urgentShortest-delayDual pricing (bounded)
Records (Table 7)13%0.5040.513
Prompted Qwen3-32B41%0.441 ± 0.0290.456 ± 0.065
Fine-tuned Qwen2.5-1.5B54%0.436 ± 0.0490.470 ± 0.010

Table R2. UTA under scaled bed capacity for policies that see SICL's inputs (five runs each).

Policy×0.8×1.0×1.2
SICL0.4460.5400.557
PPO0.3850.4390.520
Qwen3-32B + Shortest-delay0.3610.4410.456
Qwen3-32B + Dual pricing (bounded)0.3810.4560.528

We will add these results and clarifications to the paper. We respond to each reviewer in their thread.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.