acceptodds
Under review as a conference paper at ICLR 2027

Look Before You Leap: Planning with Risk and Cost Awareness for a Clinical Diagnosis Agent

Abstract

Clinical diagnosis unfolds as a sequence of consequential decisions: every test reveals information, consumes resources, and changes what should be done next. Yet diagnostic agents typically decide one step at a time, without explicitly considering how an action may shape the future diagnostic trajectory. We ask whether an agent can make better decisions by looking ahead before acting. To study this question in realistic clinical settings, we construct FISC, a Fast healthcare interoperability resources-based Interactive Sequential Clinical case dataset derived from 156 longitudinal real-world patient records, enabling standardized and interactive diagnostic trajectories. We then introduce SPARC, a Sequential Planning with Anticipated Risk and Cost agent, which formulates diagnosis as resource- and risk-aware planning under partial observability. Before executing a test, SPARC imagines its plausible outcomes and their downstream diagnostic trajectories, and uses stochastic AND–OR search to choose actions by jointly considering information gain, cumulative cost, diagnostic error, and the risk of missing critical disease. Following receding-horizon planning, SPARC executes only the selected next action and replans after observing its real outcome. Across real-world clinical cases, SPARC achieves diagnostic performance comparable to alternative sequential strategies while using fewer tests, and its behavior across planning depths reveals a trade-off between lookahead horizon, resource use, and diagnostic performance. These results suggest that diagnostic actions can be usefully evaluated not only by their immediate utility, but also by their anticipated downstream consequences.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Sequential clinical diagnosis, Multi-step planning, Risk and cost aware decision making

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_rSETOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper formulates sequential diagnostic testing as a planning problem under partial observability, on the premise that a test's value should be judged by the downstream diagnostic trajectory it enables rather than by immediate information gain. It contributes FISC, a dataset of 156 longitudinal inpatient records re-encoded as HL7 FHIR resources across ten clinical domains, with a standardized test catalog carrying codes and prices and a designated set of time-critical conditions (TCCs); SPARC, a receding-horizon planner whose State, Dreamer, and Explore modules perform finite-depth stochastic AND–OR search over imagined outcome branches, scoring actions by Q(S,a) = C(a) + Σ_x P(x|S,a) V(T(S,a,x)) and stopping under the safety constraints b(d̂) ≥ τ_B and L_extra_TCC ≤ τ_M; and an evaluation against four baselines across planning depths D0–D5 and two result environments, reporting Top-k accuracy, Top-k TCC miss rate, mean test cost, and mean test count. All experiments are run on FISC alone.

Critical strengths

1. The problem formulation is well-motivated.

Casting inpatient diagnosis as a POMDP in which an agent iteratively selects tests, observes results, and updates beliefs — then evaluating candidate actions through their anticipated downstream consequences rather than immediate information gain — is a sound and well-argued framing. Making the risk of missing a time-critical condition an explicit asymmetric terminal loss rather than absorbing it into diagnostic accuracy matches clinical reality. This matters because the formulation is the part most likely to outlive the specific results, and it is worth building on.

2. The paper contributes real clinical data and a relatively complete evaluation system.

FISC is built from 156 real longitudinal inpatient records re-encoded as HL7 FHIR, spanning ten clinical domains, with a standardized test catalog carrying codes and prices and a designated TCC set. The evaluation covers diagnostic accuracy, TCC miss rate, resource cost, test count, and stopping statistics across planning depths and two result environments, which is a more complete evaluation protocol than most work in this area. The evaluation substrate — FHIR-grounded action space, real cost table, TCC-aware metrics — is reusable and is a genuine contribution independent of whether this particular method succeeds. The paper is also clearly written and well organized, and the running example in Figure 1 makes the central motivation immediately legible.

Critical weaknesses

1. The method does not constitute a substantial advance over standard receding-horizon planning.

SPARC combines three established ingredients — a belief update, LLM-simulated test outcomes, and finite-depth search over imagined branches — under a receding-horizon scheme. The paper does not articulate what is new relative to generic active feature acquisition or Bayesian experimental design, the latter being the principle behind ACTMED, one of its own baselines. The apparatus that would carry the novelty, namely the objective Q(S,a) and the stopping constraints b(d̂) ≥ τ_B and L_extra_TCC ≤ τ_M, is presented without any analysis: there is no guarantee, no characterization of when lookahead helps, and some quantities it depends on are not fully specified. As written, the contribution reads as a careful engineering assembly rather than a methodological advance.

2. The experiments do not establish that lookahead improves diagnostic decisions.

Section 6.2 states that one-step planning improves diagnostic accuracy while reducing the TCC miss rate, testing cost, and number of tests. The reported depth sweep does not support this: accuracy across D0–D5 is 60.4, 63.5, 57.7, 59.6, 61.5, 60.9, and TCC miss rate is 41.5, 40.4, 44.2, 50.0, 46.2, 40.4, so the trend is non-monotonic and D2 and D3 are worse than D0 on both measures. Nor can the evaluation compensate: all experiments are run on FISC alone, and the LLM driving the State, Dreamer, and Explore modules is never identified, so the numbers cannot be separated from a particular unnamed model and cross-model or cross-dataset generalization cannot be assessed. Since the title, abstract, and RQ2 rest on this claim, the central proposition is left unsupported.

Taken together, the formulation and the evaluation substrate are worth keeping, but the method is a standard receding-horizon planner and the experiments do not show that it improves diagnostic decisions significantly.

Questions for authors
  1. Model and evaluation scope. Which LLM drives the State, Dreamer, and Explore modules — model family, version, and scale? Given that the paper reports no model identity and evaluates on FISC alone, can the authors report results for at least one additional backbone and, if available, on a second dataset, so that the findings can be separated from the specific model and data used?

  2. Reconciling §6.2 with Appendix Table 2. Table 2 reports paired 95% CIs of +3.0 [−2.2, +8.3] for accuracy, −1.1 [−10.0, +7.8] for TCC miss rate, and −6 [−25, +13] CNY for cost — all containing zero — while §6.2 states these as improvements. Could the authors reconcile the main-text claim with their own paired analysis and state plainly which effects the experiments establish?

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

This paper introduces SPARC, a clinical diagnosis agent that incorporates risk- and cost-aware lookahead planning into sequential diagnostic decision-making. The proposed framework uses stochastic AND–OR search to anticipate possible test outcomes and select diagnostic actions while considering testing costs and the risk of missing time-critical conditions. The authors also construct FISC, an interactive clinical dataset based on 156 real-world patient records, and evaluate SPARC against several sequential diagnostic strategies in terms of diagnostic accuracy, testing costs, and resource utilization.

Critical strengths
  • The paper addresses a clinically relevant problem by jointly considering diagnostic accuracy, testing costs, and the risk of missing time-critical conditions.

  • The proposed framework integrates lookahead planning and risk-aware decision-making into a structured sequential diagnosis workflow.

  • The FISC dataset provides an interactive evaluation environment based on real-world clinical records and standardized FHIR representations.

Critical weaknesses
  • The proposed method lacks novelty. This work highly depends on modifying the conversational strategies with an agent. The methodology does not involve updating/fine-tuning agent parameters or accumulating external knowledge.
  • Experimental settings are incomplete. The authors have not yet provided any details on the adopted LLM, its version, and the number parameters. Therefore, it is unclear if the proposed strategy can really outperform SOTA LLMs. Furthermore, it is unclear if the proposed approach is still effective when applied to SOTA LLMs. Overall, no empirical evidence can show that current SOTA LLMs still exists the claimed issues in this paper.
  • Insufficient experimental validation. First, the critique of the "greedy" approach is questionable. With a SOTA LLM, a greedy strategy might achieve an optimal solution. Besides, the highest accuracy is only observed at D1 rather than at deeper levels (D2–D5), which is insufficient to demonstrate the effectiveness of the "lookahead" mechanism. This work also lacks ablation studies for important components and generalization studies across centers.
Questions for authors
  • Could the authors provide comparisons with stronger reasoning-capable LLMs or carefully designed prompting baselines to better establish the necessity of the proposed planning framework?

  • Why does increasing the planning depth beyond D1 fail to consistently improve diagnostic performance? Could the authors provide further analysis of the potential causes?

  • Could the authors provide ablation studies for the key components of SPARC and generalization studies across centers?

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

This paper proposes SPARC, a risk- and cost-aware clinical diagnosis agent that uses stochastic AND-OR lookahead and receding-horizon replanning to anticipate the downstream consequences of diagnostic tests. It also introduces FISC, a FHIR-based interactive dataset derived from 156 longitudinal patient records. Experiments report diagnostic performance comparable to alternative sequential strategies with fewer tests, while revealing non-monotonic trade-offs among planning depth, cost, and safety.

Critical strengths
  • The paper addresses the clinically important problem of balancing diagnostic accuracy, resource consumption, and the asymmetric harm of missing time-critical conditions.
  • SPARC combines counterfactual outcome prediction, stochastic AND-OR search, safety-constrained stopping, and FHIR-compatible actions in a coherent sequential planning framework.
  • The controlled comparison across planning depths, paired D0-D1 analysis, and evaluation under both recorded-only and simulated-result environments provide several useful views of the system's behavior.
Critical weaknesses
  • The evaluation uses only 156 retrospective cases from one institution, including roughly 32 time-critical cases, while the private dataset prevents independent reproduction and substantially limits evidence for generalization.
  • One-step lookahead significantly reduces the average number of tests relative to D0, but its reported improvements in Top-1 accuracy, TCC miss rate, and testing cost are not statistically established because their paired 95% confidence intervals include zero; moreover, the cross-method comparisons evaluate different overall decision strategies and therefore cannot isolate the effect of lookahead.
  • The simulated-result evaluation has serious validity and reproducibility gaps: Appendix H.7 claims a fidelity study but reports no sample size or numerical results, Appendix F.5 instead states that fidelity was not evaluated, and the exact language model, inference settings, and proposed component-ablation results are not reported.
Questions for authors

See Weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.