acceptodds
Under review as a conference paper at ICLR 2027

CARE: A Consulting, Acquiring, Resource-aware, Explaining Agent for Clinical Decision Support

11 views

Abstract

Clinical decision support is based on a record in progress: tests are ordered as the diagnosis evolves, each with medical resource consumption, and the diagnosis process is personal per patient condition. We present CARE, a generic framework for this setting. A reinforcement learning agent for predicting medical outcome, trained on the medical center's own patients, works in a human-like process, deciding per patient which test to order next, when to consult a medical language model, and when to stop and predict, under test-specific costs and turnaround times and within resource budget limits. The language model supplies medical knowledge and a learned fusion model weighs its assessment by the evidence behind it and fuses it into the locally trained prediction. Every episode is rendered into a clinical account of what was acquired and what it showed, so a clinician can see how the recommendation was reached. On MIMIC-IV, CARE reaches higher accuracy than dynamic acquisition agents and static test rankings at an earlier diagnosis stage, while keeping its workups individual to the patient. The framework is generic and can be applied to any medical data and any medical outcome, a holistic approach that combines local learning, medical knowledge, resource awareness, and self-explanation in one system.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
clinical decision support, active feature acquisition, MIMIC-IV, large language models, reinforcement learning, semi-Markov decision process, explainability, electronic health records, cost-sensitive learning

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_obBnOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2(+1)
Presentation
1
Soundness
2(+1)
Read review
Paper summary

This paper presents CARE, a generic framework that meets the four properties of clinical decision support. Clinical features are represented through a shared text encoder, while acquisition costs and durations influence rewards and stopping decisions. Experiments on MIMIC-IV report improved mortality discrimination with fewer acquired panels, alongside separate evaluations of resource awareness and explanation quality.

Critical strengths
  1. CARE provides a framework for adaptive clinical evidence acquisition, combining test selection, optional language-model consultation, learned prediction fusion, and explicit resource constraints. Treating consultation and stopping as learned actions is a practically relevant design.

  2. The reported results show improved predictive efficiency within the benchmark: CARE achieves a mortality AUROC of 0.9240 with fewer mean features acquired than EDDI. Full-record fusion also outperforms both constituent predictors, supporting their predictive complementarity.

Critical weaknesses
  1. The agent may see test results before they would be available in real clinical practice. A panel can include measurements taken up to 24 hours after admission, but the agent receives the whole panel after a fixed waiting time (Appendices B and J). The paper does not show that every result would already be available then. The agent may therefore use later information to make an apparently earlier prediction, so selecting fewer panels does not establish earlier clinical decision-making.

  2. The methodological novelty is limited and insufficiently validated. CARE largely combines established prediction and reinforcement-learning components, while its distinctive design is learning when to consult the LLM. However, the consultation ablation only shows that adding LLM predictions helps. It does not establish that learning the consultation timing is necessary. The paper should test whether learning when to consult the LLM works better than simply consulting it at a fixed step, while keeping the models and budget unchanged.

Questions for authors

See Weakness.

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

This paper introduces CARE, a retrospective clinical decision-support framework for sequentially selecting EHR test panels, optionally consulting MedGemma, and stopping to predict in-hospital mortality and 30-day readmission. A local reward model predicts outcomes from acquired features; a fusion model combines its prediction with the LLM assessment; and a dueling double-DQN policy selects panels under cost and turnaround-time constraints, with a semi-Markov formulation for duration. On MIMIC-IV, CARE reports higher mortality AUROC at fewer acquired panels than EDDI, static SHAP-based rankings, and SM-DDPO.

Critical strengths
  1. The paper tackles a genuinely important formulation that is more realistic than fixed-feature clinical prediction: evidence is acquired sequentially, varies by patient, and has resource constraints. Treating LLM consultation as an explicit action and fusing it with a locally trained model is a thoughtful design that avoids relying on the LLM alone for prediction.

  2. The empirical study is unusually broad for this setting. It includes dynamic and static baselines, a no-consultation ablation, cost/time tradeoffs, and an attempt to evaluate explanation faithfulness. The result that CARE reaches the reported AUROC of 0.9240±0.00180.9240 \pm 0.0018 with about 3.6 panels, whereas EDDI requires substantially more panels to reach a similar level, is promising.

Critical weaknesses
  1. The central clinical/resource claims are not adequately validated by the retrospective setup. The policy is evaluated on panels and results that were generated under clinicians’ historical ordering behavior; setting absent panels to zero does not establish what would have happened had CARE ordered an unobserved test. Moreover, panels are constructed from broad admission-relative windows, while turnaround is represented by population-level median duration and acquisitions are sequential. This makes it unclear whether the reported gains in “tests,” cost, and especially elapsed time would hold for a deployable ordering policy, where tests can be ordered in parallel and test availability depends on clinical state and workflow. A temporally held-out or external evaluation with strictly action-available timestamps, together with a more defensible off-policy/counterfactual treatment, is needed for the paper’s main practical claims.

  2. Evaluation is narrow for a high-stakes prediction system. Although the formulation includes mortality and readmission, the principal comparisons are mortality AUROC only. The paper does not report calibration, subgroup performance, uncertainty/statistical testing, or downstream clinical utility; several results use only two or three random seeds. The explanation evaluation also relies on an LLM judge and is reported for the agent without consultation, rather than clinician assessment of whether explanations are useful, safe, and appropriately tied to the actual recommendation. These omissions leave the claimed clinical usefulness and “self-explanatory” property insufficiently supported.

Questions for authors
  1. Please provide a precise timestamp audit: for every panel, what timestamp determines its availability to the policy, and how do you ensure no result, note content, or derived representation contains information unavailable when the panel would be ordered? Can you rerun the main evaluation using strictly prospective availability relative to admission and prior actions?

  2. How should the retrospective policy value be interpreted when a panel not historically ordered is represented as absent/zero? Do you have an off-policy evaluation, an analysis restricted to clinically feasible actions, or evidence that the policy’s selected panels would have been available under intervention?

  3. Can you report mortality and readmission calibration, uncertainty/confidence intervals and statistical comparisons, plus clinically meaningful subgroup analyses? For the explanation claim, can you evaluate consultation-enabled CARE with blinded clinician raters or otherwise justify why the LLM-judge score represents clinical usefulness rather than trace consistency?

Reviewer_zaQ2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CARE, an agent that sequentially acquires patient information, optionally consults a medical language model, and predicts clinical outcomes under resource constraints. It also generates explanations of the acquisition process.

Critical strengths

CARE performs better with consultation than without it, and the full-record fusion model also outperforms either predictor alone. These results suggest that the language model provides complementary predictive information in this setting.

Critical weaknesses
  • Fewer tests do not necessarily mean a more appropriate clinical workup. Some tests are needed to rule out differential diagnoses or follow clinical guidelines, even when the results are negative or the patient has no related symptoms. CARE optimizes outcome prediction under resource constraints, but a test that contributes little to mortality prediction may still be important for diagnosis or management. How does the agent distinguish an unnecessary test from a necessary rule-out test? Please evaluate whether the omitted tests and stopping decisions are clinically appropriate, through guideline-based assessment or clinician review, rather than interpreting fewer acquired panels alone as better clinical decision-making.

  • The definition of a panel makes the resource savings difficult to interpret. For example, "routine chemistry" includes renal and liver measurements, lipids, HbA1c, vitamins, and inflammation markers, but counts as one acquisition in the lowest cost tier. This is a broad collection of information, not clearly one orderable test. Duration is also represented through coarse time units, with imaging durations assigned by modality and acquisition times added sequentially.

  • The complete framework is not evaluated as a complete system. The main contribution is presented as combining consultation, resource awareness, and explanation. However, Table 2 reports explanation results for the agent without consultation, and Figure 5 evaluates resource awareness without consultation. These separate experiments do not establish that all the benefits hold together.

  • The baseline comparisons are not sufficient to establish the advantage over closely related methods. The paper includes EDDI, SM-DDPO, and static rankings, but several relevant medical acquisition methods discussed in the related work are not evaluated. e.g, ED-Copilot (Sun et al. 2024) already combines language-model representations with patient-specific, time-aware test selection, while P-CAFE (Kashani et al., 2025) addresses personalized, cost-aware acquisition from EHRs. These seem particularly relevant to CARE. Please include comparisons under comparable feature sets and resource budgets, or explain why these methods cannot be meaningfully compared. Without these comparisons, it is difficult to determine what CARE adds beyond existing approaches.

  • The generated account describes acquired tests and gives plausible clinical interpretations. However, a clinically reasonable explanation written afterward is not necessarily the reason the policy selected those tests. Checking values against the trace and pairing traces with mismatched accounts mainly tests factual consistency and patient specificity. It does not establish faithfulness to the acquisition policy. Please distinguish a trajectory summary from an explanation of the agent’s decision process, and include clinician assessment of the explanations’ usefulness and correctness.

  • The clinical claims are broader than the reported evaluation. The experiments mainly evaluate mortality AUROC on one dataset. Predicting mortality risk is not the same as identifying a diagnosis or establishing that further clinical testing is unnecessary. Readmission is included in the model and reward formulation, but its performance is not reported. Please report that outcome, calibration, and performance at clinically relevant thresholds. Claims that the framework applies to any medical outcome also need qualification without evaluation on additional tasks or settings.

Questions for authors
  • minor suggestion: Use AUROC rather than accuracy when describing those results. Also distinguish fewer acquired panels from fewer individual tests or shorter actual clinical waiting time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.