acceptodds
Under review as a conference paper at ICLR 2027

Complementary Evidence Reasoning over Structured and Unstructured Electronic Health Records

Abstract

Structured and unstructured features derived from the same record often overlap, so learning from both requires distinguishing complementary evidence from redundant information. In Electronic Health Records (EHRs), clinical notes capture key observations that are not readily standardized into medical codes, such as symptom severity and treatment response, while much of their remaining content is already in the structured record. Existing EHR methods encode or summarize notes before fusing them with the structured record, without asking what the note adds beyond the codes. Reasoning-enhanced methods reason to represent an input on its own terms, without feedback from the downstream task. In both cases, reasoning is not trained to select complementary evidence for predictive utility. To this end, we propose CERE (Complementary Evidence Reasoning for EHRs), which uses LLM reasoning to learn task-aligned representations of clinical notes that complement structured records. CERE first maps visit codes to soft prompts that tell the LLM what the structured record already captures. Guided by these prompts, the LLM reasons about what the note adds and encodes that reasoning as a dense representation trained for prediction. Finally, reinforcement learning refines the reasoning itself, rewarding evidence selections that lower prediction loss. Experiments on MIMIC-III and MIMIC-IV show that CERE consistently outperforms strong baselines across multiple clinical prediction tasks.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
electronic health records, large language models, LLM reasoning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_QE7vOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CERE (Complementary Evidence Reasoning for EHRs), a framework for integrating structured electronic health records (EHRs) and unstructured clinical notes by identifying complementary information that is not already captured by structured medical codes. The framework consists of three optimization stages: (1) structured context grounding, which converts diagnosis, procedure, and medication codes into continuous soft prompts using a discriminative reconstruction objective; (2) residual evidence learning, which employs an LLM to generate code-conditioned clinical reasoning and extracts task-specific representations through predictive LoRA adaptation; and (3) predictive-guided reasoning refinement, which uses group-relative reinforcement learning to optimize reasoning trajectories based on downstream prediction performance. The resulting representations are integrated with structured EHR features through a RETAIN-based predictor. Experiments on MIMIC-III and MIMIC-IV evaluate the framework on hospital readmission prediction and medication recommendation, demonstrating improvements over conventional EHR models, LLM-based approaches, and frozen clinical note encoders.

Critical strengths

S1. The paper addresses an important limitation of existing multimodal EHR models: the potential redundancy between structured medical codes and clinical notes. Learning complementary rather than merely combined representations is a meaningful research direction.

S2. The three-stage framework establishes a logical progression from structured context grounding to task-specific representation learning and reinforcement learning-based reasoning refinement. Each stage addresses a distinct aspect of the proposed problem.

S3. Using downstream prediction loss to guide reasoning refinement is an interesting approach. It provides a mechanism for connecting LLM-generated reasoning with predictive utility rather than relying solely on linguistic plausibility.

S4. The appendices provide architectural specifications, training procedures, prompts, hyperparameters, and reinforcement learning details, improving the reproducibility of the proposed method.

Critical weaknesses

W1. Although the central claim is that CERE identifies information absent from structured records, the experiments primarily measure downstream prediction performance. These improvements do not necessarily demonstrate that the learned representations specifically capture complementary rather than redundant information. A more direct evaluation of evidence complementarity would strengthen the paper.

W2. CERE requires teacher-generated descriptions, LLM reasoning generation, multiple optimization stages, patient-level RL rollouts, and a final predictor refit. The paper does not provide a sufficiently detailed comparison of computational costs, inference latency, or performance-efficiency trade-offs against simpler fusion methods.

W3. The experiments use two MIMIC datasets originating from the same medical center, with MIMIC-IV further restricted to 20% of eligible patients. Consequently, the results do not establish robustness across different healthcare institutions, documentation practices, or patient populations.

W4. The case studies illustrate changes in evidence prioritization after RL, but they do not independently establish that the selected evidence is clinically accurate, genuinely absent from structured records, or consistently useful across patients.

Questions for authors

Q1. The paper defines the residual representation as z^{res}=F(S\mid V). However, this formulation does not mathematically enforce the removal of redundant information or the isolation of complementary information. The representation is conditioned on structured codes and optimized for prediction, but it may still encode information already present in those codes. The authors should clarify whether complementarity is an intended inductive bias or a property explicitly enforced by the learning objective.

Q2. During Stage 3, reasoning trajectories are rewarded using a frozen predictor trained on Stage-2 representations. After RL, the predictor is reinitialized and retrained from scratch. Consequently, the reward optimizes performance under a predictor that differs from the final evaluation model. Although this procedure is explicitly described, the paper does not sufficiently investigate whether the learned evidence-selection policy remains optimal after the final predictor refit.

Q3. The authors conclude that RL provides additional benefits, which is supported by the higher mean scores. However, the reported confidence intervals overlap for several comparisons between CERE and its variant without RL. Overlapping confidence intervals do not establish a lack of significance, but paired statistical tests would be needed to support stronger claims about the reliability of the improvements.

Q4. The case study suggests that RL shifts attention toward clinically relevant risk factors. However, the selected factors are not systematically compared against the structured codes to verify that they represent genuinely complementary evidence. The appendix appropriately acknowledges this limitation for its additional case study. A quantitative assessment of evidence novelty and factual consistency would better support the paper’s central hypothesis.

Q5. The RL reward combines prediction loss, formatting quality, and representation drift. Since the same adapters influence both generated reasoning and extracted representations, the policy may improve the frozen predictor’s loss without necessarily improving the semantic quality of evidence selection. The authors acknowledge that the regularization terms do not guarantee the absence of reward exploitation. Nevertheless, the paper would benefit from an empirical analysis separating improvements in reasoning quality from changes in embedding geometry.

Q6. The frozen note-encoder baselines provide useful comparisons, but CERE additionally benefits from task-specific LLM adaptation and reinforcement learning. The existing ablations partially address these differences, yet a comparison with a similarly adapted note encoder without complementary reasoning would help isolate the contribution of the proposed reasoning strategy.

Q7. he paper defines the medication prediction target as prescriptions at the next visit and uses historical visits as inputs. However, the clinical interpretation of this setup would benefit from further clarification, particularly regarding the prediction time, availability of discharge notes, and whether all included information would be accessible when recommendations are made.

Q8. The baseline description refers to RePrompT as integrating structured longitudinal EHR encoders with “doctoral notes”. Based on the surrounding discussion, this appears to be a typographical error for “clinical notes” or a similar term.

5 comments
AuthorsOpenReview ↗

Response to Reviewer QE7v

Thank you reviewer QE7v, for the detailed assessment and for raising questions and concerns about our research. We address them below:

W1, W4, Q1, and Q4: Complementarity and the meaning of “residual.”

Complementarity is an intended inductive bias, rather than a mathematically guaranteed exclusion of redundant information. Section 2.1 explicitly uses “residual” to describe the conditional predictive role of the note representation. The objective is to extract useful evidence from notes given the structured record, rather than enforce independence between the two representations. Structured-code grounding, code-conditioned reasoning, and prediction jointly with structured inputs implement this objective by providing the coded context and helping the model to identify additional evidence relevant to the target.

The relevant comparisons from Table 3 are reproduced below. Removing structured conditioning or generated reasoning lowers mean performance across all eight dataset–metric combinations. The variant without residual reasoning retains structured conditioning, predictive LoRA adaptation, and the downstream predictor, providing a task-adapted direct-encoding comparison.

DatasetModel variantReadmission AUROCReadmission PRAUCMedication PRAUCMedication F1
MIMIC-IIIWithout conditional encoding73.0537.2134.3219.11
MIMIC-IIIWithout residual reasoning74.1637.5335.2420.43
MIMIC-IIICERE75.1338.4336.3322.48
MIMIC-IVWithout conditional encoding66.1337.9037.9222.74
MIMIC-IVWithout residual reasoning67.8938.4838.7324.23
MIMIC-IVCERE68.7139.4339.5425.11

We further assessed the generated evidence directly on 3,000 randomly selected held-out cases. Clinical agent evaluators, blinded to model-variant identities, reviewed the original notes, supplied codes and their descriptions, and generated outputs. Each output received a score between 0 to 1 for factual accuracy, information beyond the supplied codes, and task relevance. CERE achieved the highest scores across all three criteria:

Model variantFactual accuracyInformation beyond codesTask relevance
Without structured conditioning0.740.620.87
Before RL refinement0.800.710.88
CERE0.870.760.93

W2: Computational cost and inference latency.

Thank you for raising this point. We prioritize predictive accuracy over computational efficiency in medical applications because prediction errors can have serious clinical consequences. For transparency, we report both training and inference times. The tables shows MIMIC-III readmission task training/inference time.

ModelTraining timeInference per patient
RETAIN0.23 h84 ms
Deepr0.20 h74 ms
GRAM0.73 h266 ms
Transformer0.18 h64 ms
AdaCare0.20 h70 ms
StageNet0.70 h252 ms
GRASP0.83 h525 ms
ADORE0.28 h112 ms
ARCI0.67 h970 ms
LINKO1.27 h490 ms
RePrompT9.52 h3.5 s
BERT + RETAIN0.44 h61 ms
ModernBERT + RETAIN0.65 h151 ms
BioClinical-ModernBERT + RETAIN1.02 h219 ms
Qwen3-Embedding-4B + RETAIN4.12 h680 ms
CERE55 h8.2 s

W3: Dataset scope.

Although both datasets originate from the same institution, this does not imply a clinically homogeneous patient population. The official MIMIC-III documentation explicitly describes a diverse ICU population, and the original dataset paper reports patients across coronary care, cardiac surgery recovery, medical, surgical, and trauma/surgical intensive care units [1]. This breadth supports its value as a clinically heterogeneous benchmark for evaluating our method.

[1] Scientific Data (2016) MIMIC-III, a freely accessible critical care database.

Our choice of MIMIC-III and MIMIC-IV follows established evaluation practice in prior work, including GraphCare [2] and KARE [3], both of which evaluate their EHR prediction methods exclusively on these two datasets.

[2] ICLR (2024) GraphCare: Enhancing Healthcare Predictions with Personalized Knowledge Graphs.

[3] ICLR (2025) KARE: Reasoning-Enhanced Healthcare Predictions with Knowledge Graph Community Retrieval.

AuthorsOpenReview ↗

Response to Reviewer QE7v - Part 2

Q2: Transfer from the frozen reward predictor to the refitted predictor.

The frozen Stage-2 predictor provides a fixed evaluation function during policy refinement, and the final refit adapts the downstream predictor to the refined representations. Importantly, Table 3 already evaluates whether the refinement benefits persist after refitting: Appendix A.5 specifies that both CERE and the variant without RL regenerate their representations and retrain a newly initialized predictor using identical settings. The table below reports the differences calculated by subtracting the “w/o RL Refinement” row from the “CERE” row in Table 3. CERE achieves higher mean performance across all eight comparisons, supporting transfer of the refinement benefit to a separately fitted predictor of the same architecture.

DatasetReadmission AUROCReadmission PRAUCMedication PRAUCMedication F1
MIMIC-III+0.51+0.34+0.49+0.51
MIMIC-IV+0.63+0.56+0.79+1.15

Q3: Reliability of RL improvements.


We propose a paired permutation test comparing CERE with its variant without RL for each dataset, task, and metric, using the mean difference across six matched runs. We randomly swap predictions between models at the patient level, keeping visits and labels together and swaps consistent across runs. We recompute this difference over 10,000 permutations and calculate two-sided p-values with a plus-one correction. The maximum and minimum of p-values are [0.0021, 0.0028] for AUROC, [0.0029, 0.0035] for PRUAC, [0.0036, 0.0045] for F1, shows statistically significant different between pairs across all seeds. We will update the paper to reflect this.

Q5: Reasoning quality versus representation changes.

Stage 3 jointly improves evidence generation and its representation for downstream prediction; changes in embedding geometry are therefore part of the intended optimization. To examine whether refinement also improves generated content, we conducted the additional held-out evidence evaluation reported in our response to W1/W4/Q1/Q4. Also, Table 3 shows higher mean predictive performance after retraining a newly initialized predictor, which supports benefits that persist beyond evaluation by the frozen reward predictor.

Q6: Task-adapted note-encoder baselines.

Thank you for this suggestion. We used frozen note encoders to measure the benefit of adding pretrained text representations without task-specific encoder training or explicit reasoning. Our evaluation also includes a task-adapted no-reasoning control: the “w/o Residual Reasoning” variant in Table 3, described in Section 3.5, retains CERE’s backbone, structured conditioning, predictive LoRA adaptation, and downstream predictor while extracting representations without generated reasoning.

Following your suggestion, we have additionally fine-tuned the baseline text encoders for each prediction task, updating the encoder and predictor using the downstream prediction loss while keeping the structured inputs, fusion architecture, and data splits unchanged. The results below show that CERE maintains higher performance across both datasets and all reported metrics, extending the comparison beyond frozen encoders.

MIMIC-III

ModelReadmission AUROCReadmission PRAUCMedication PRAUCMedication F1
BERT - tuned73.3936.4735.0920.66
ModernBERT - tuned73.7836.7335.2721.07
BioClinical-MBERT - tuned74.0337.1235.4621.43
Qwen3-Embedding-4B - tuned74.2837.4935.7321.89
CERE75.1338.4336.3322.48

MIMIC-IV

ModelReadmission AUROCReadmission PRAUCMedication PRAUCMedication F1
BERT - tuned65.8336.1237.7922.23
ModernBERT - tuned66.2636.7438.1322.61
BioClinical-MBERT - tuned66.8937.3238.4123.08
Qwen3-Embedding-4B - tuned67.4738.1938.8624.03
CERE68.7139.4339.5425.11

Q7: Prediction time and discharge-note availability.

As defined in Sections 2.1 and 3.2, we predict medications at admission (T+1) using completed admissions through (T), after their documentation is available. Historical medication records and discharge notes are therefore valid inputs, and medications and notes from the target admission are excluded.

Q8: Typographical correction.

Thank you for identifying this error; we will correct it in the final version.

AuthorsOpenReview ↗

Follow-up on our responses

Dear Reviewer QE7v, Before the discussion closes, we would appreciate hearing your thoughts on our responses and additional results. If there are concerns we have not fully addressed, please let us know so we can respond. Thank you for your time and feedback.

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

This paper studies complementary evidence learning between structured EHR codes and unstructured clinical notes. Existing methods either encode or summarize notes before fusing them with codes, or reason over the note on its own terms—none train reasoning to select evidence that adds predictive value beyond the codes. The authors propose CERE: first, visit codes are mapped to soft prompts via a discriminative reconstruction objective, telling the LLM what the structured record already captures; second, guided by these prompts, the LLM reasons about what the note adds, and the <embed> hidden state provides a residual representation fused with structured features in a RETAIN-style predictor, adapted end-to-end through the prediction loss via LoRA; third, group-relative GRPO refines evidence selection in the reasoning trajectories, using the frozen predictor as judge and rewards combining format quality, prediction loss, and representation drift. On MIMIC-III and MIMIC-IV, CERE achieves the best results on 15-day readmission prediction and medication recommendation, outperforming 9 deep-learning baselines, 3 LLM-based methods, and 4 frozen-encoder fusions; ablations show the largest drops when conditional encoding or grounding contrastive loss is removed, and a case study shows reasoning shifting from measurement details to readmission-relevant factors after GRPO.

Critical strengths

S1: The three-stage pipeline (grounding → residual evidence learning → predictor-guided refinement) has a clear dependency ordering and a sensible sequential optimization.

S2: Well-designed RL rewards: group-relative normalization handles cross-patient difficulty, and the reward combines format quality, prediction loss, and a representation-drift penalty.

S3: Component ablations are comprehensive and consistent across datasets and tasks, with the largest drops from removing conditional encoding and the grounding contrastive loss.

Critical weaknesses

W1: Modest gains over the strongest baseline: about +1.9/+2.2 AUROC over a frozen encoder with no task adaptation, and no comparison against a task-adapted text encoder.

W2: Computational cost is high and unreported: per-visit LLM reasoning plus G rollouts per patient, with no cost, token, or latency comparison against baselines.

W3: The judge-predictor loop may bias rewards: the frozen Stage-2 predictor judges rollouts while the final predictor is retrained from scratch on refined representations, and key thresholds (e.g., τ_d) lack sensitivity analysis.

Questions for authors

See weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer RVye

Thank you Reviewer RVye, for your helpful comments. We address your concerns below.

W1: Performance Gains and Task-Adapted Baselines

Thank you for this concern. We used frozen note encoders to assess how much improvement comes from pretrained text representations alone, without additional encoder adaptation. This provides a reference for evaluating the benefit of CERE’s task-specific training and reasoning. We have also evaluated task-adapted versions of the same encoders, updating the text encoder and predictor using the downstream prediction loss and keeping the structured inputs, fusion architecture, and data splits unchanged:

MIMIC-III

ModelReadmission AUROCReadmission PRAUCMedication PRAUCMedication F1
BERT73.3936.4735.0920.66
ModernBERT73.7836.7335.2721.07
BioClinical-MBERT74.0337.1235.4621.43
Qwen3-Embedding-4B74.2837.4935.7321.89
CERE75.1338.4336.3322.48

MIMIC-IV

ModelReadmission AUROCReadmission PRAUCMedication PRAUCMedication F1
BERT65.8336.1237.7922.23
ModernBERT66.2636.7438.1322.61
BioClinical-MBERT66.8937.3238.4123.08
Qwen3-Embedding-4B67.4738.1938.8624.03
CERE68.7139.4339.5425.11

Our method consistently outperforms these tuned encoders across all tasks.

W2: Computational Cost and Inference Latency

Thank you for raising this point. We give more importance to predictive accuracy in medical applications because errors can have serious consequences. Computational efficiency also matters, so we report both training and inference times.

MIMIC-III readmission: training and per-patient inference

ModelTraining timeInference per patient
RETAIN0.23 h84 ms
Deepr0.20 h74 ms
GRAM0.73 h266 ms
Transformer0.18 h64 ms
AdaCare0.20 h70 ms
StageNet0.70 h252 ms
GRASP0.83 h525 ms
ADORE0.28 h112 ms
ARCI0.67 h970 ms
LINKO1.27 h490 ms
RePrompT9.52 h3.5 s
BERT0.44 h61 ms
ModernBERT0.65 h151 ms
BioClinical-ModernBERT1.02 h219 ms
Qwen3-Embedding-4B4.12 h680 ms
CERE55 h8.2 s

One GRPO step, including G=8G = 8 rollouts for one patient and the policy update, takes 185 s, with an average of 1,923 tokens generated across the eight rollouts.

W3: Reward Bias and Threshold Sensitivity

Thank you for this comment. In our no-RL ablation, we retrain the predictor from scratch with the same settings as CERE, as stated in Appendix A.5. Table 3 therefore shows that the gains remain after retraining the predictor. We also tested whether these gains depend on the judge by training a Transformer predictor on representations with and without RL. RETAIN supplied all RL rewards; the Transformer was trained separately and never used as a judge. The Transformer also improved with RL, suggesting that the gains are not limited to the RETAIN judge.

Final predictorRepresentationsMIMIC-III readmission AUROCPRAUC
RETAIN + RETAINWithout RL74.6238.09
RETAIN + RETAINWith RL75.1338.43
RETAIN + TransformerWithout RL74.3137.63
RETAIN + TransformerWith RL74.7437.94

We have evaluated sensitivity to the drift tolerance, drift penalty, and reference-policy penalty on MIMIC-III readmission. Each setting was varied while keeping the remaining hyperparameters fixed.

Setting variedValueAUROCPRAUC
Default settingsτ_d = 1.1, δ = 0.5, β = 0.0275.1338.43
Drift toleranceτ_d = 0.574.8338.22
Drift toleranceτ_d = 1.575.0338.34
Drift penaltyδ = 0.2575.0238.31
Drift penaltyδ = 1.074.9138.24
Reference-policy penaltyβ = 0.0174.9838.23
Reference-policy penaltyβ = 0.0474.9338.32

The following table reports sensitivity to the rollout group size and number of policy updates on the same task.

Rollout group sizePolicy updatesAUROCPRAUC
420074.8738.21
820075.1338.43
1620075.0838.39
810074.8138.17
840075.2438.59
AuthorsOpenReview ↗

Follow-up on our responses

Dear Reviewer RVye, With the discussion deadline approaching, we wanted to follow up on our rebuttal. Could you let us know whether we have addressed your main concerns? We are happy to clarify any remaining points. Thank you again for reviewing our work.

Reviewer_RVyeOpenReview ↗

Maintain my original score

I have read the authors' response and acknowledge that it addresses my questions. However, it does not change my evaluation, and I maintain my original score.

Reviewer_qUP8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces CERE (Complementary Evidence Reasoning for EHRs), a framework that learns task-aligned clinical note representations to capture predictive information complementary to structured electronic health records (EHRs). CERE operates in three stages: first, mapping structured clinical codes to continuous soft prompts using teacher-generated descriptions; second, conditioning a large language model on these prompts to derive note representations that are fused with structured features for downstream prediction; and third, refining the model's reasoning via group-relative policy optimization with feedback from a frozen prediction model, constrained by format and representation-drift rewards. Evaluated on MIMIC-III and a 20% subset of MIMIC-IV for 15-day readmission prediction and next-visit medication recommendation across six random seeds, CERE demonstrates superior performance over conventional EHR baselines, LLM-based approaches, and frozen note encoders, with ablation studies confirming the specific contributions of each framework component.

Critical strengths
  1. Meaningful Problem Formulation and Coherent Design: The paper addresses a key multimodal EHR challenge: standard fusion often fails to capture textual insights complementary to structured codes due to information overlap. CERE solves this by conditioning note reasoning on learned code representations and optimizing them for downstream prediction. Its three-stage pipeline (structured grounding, predictive representation learning, and policy refinement) provides a logical training framework. Using a discriminative reconstruction objective with downstream feedback effectively aligns clinical reasoning with task utility.

  2. Comprehensive Empirical Validation and In-Depth Ablation: The evaluation is rigorous, testing CERE across two datasets, two clinical tasks, and six random seeds against diverse structured, LLM-based, and frozen-encoder baselines. Extensive ablation studies isolate the contributions of structured conditioning, contrastive grounding, explicit reasoning, predictive LoRA adaptation, and reinforcement learning. Furthermore, providing anonymized code supports reproducibility and reinforces CERE's effectiveness in integrating structured and unstructured clinical data.

Critical weaknesses
  1. Insufficient Direct Validation of the Complementary-Evidence Mechanism: The paper's core claim is that CERE identifies prediction-relevant evidence in clinical notes that is genuinely complementary to structured codes. While downstream performance improvements provide indirect support, they do not directly prove that the model selects non-redundant, clinically valid evidence rather than exploiting other predictive text patterns. Furthermore, the qualitative case studies lack independent clinical evaluation to verify whether highlighted information is absent from structured inputs, an issue explicitly acknowledged in Appendix A.9. A targeted analysis assessing the clinical validity, non-redundancy, and task-specific utility of the extracted evidence is necessary to isolate the gains of complementary reasoning from general text-representation adaptation.

  2. Limited Validation and Sensitivity Analysis of Policy Refinement: Stage 3 uses predictor-guided policy refinement with format and drift constraints to optimize reasoning. However, the reward mechanism may favor trajectories that exploit the frozen predictor's specific decision boundaries rather than selecting clinically valid evidence, a risk recognized in Appendix A.3. Although ablation results show performance gains from reinforcement learning, the paper lacks sufficient analysis regarding policy stability and sensitivity to reward weights, rollout group size, or optimization budget. Clarifying patient coverage during the reported 200 policy updates is critical, and a systematic evaluation linking reward optimization to independently measured evidence quality would substantially strengthen the methodological claims.

Questions for authors
  1. Beyond downstream prediction metrics and qualitative case studies, can the authors provide an independent evaluation (e.g., via a targeted audit) to verify whether the selected evidence is clinically correct, genuinely missing from structured inputs, and improved in quality after policy refinement?
  2. Could the authors clarify the exact number of unique training patients involved in Stage 3 policy updates, the sensitivity of the performance to reward weights, rollout group size, and training budget, and whether higher predictor-based rewards directly correspond to better evidence quality rather than model exploitation?
  3. What is the precise construction and temporal alignment of prediction instances, what is the rationale and impact on label distribution for excluding single-admission patients, and how are medication-label vocabularies constructed to prevent potential data leakage?
  4. How are the reported 95% confidence intervals computed, how is the F1 score threshold for medication recommendation selected, were all baselines tuned under comparable compute and hyperparameter search budgets, and what is the exact sampling procedure used for the 20% MIMIC-IV subset?
4 comments
AuthorsOpenReview ↗

Response to reviewer qUP8

We thank Reviewer qUP8 for the careful review and constructive feedback. We address your concerns and questions below, with additional experiments and clarifications.

W1 Direct Evaluation of Complementary Evidence

Thank you for this comment. We evaluated 3000 randomly selected held-out cases in test set with clinical agent evaluators. They reviewed the original notes, the supplied codes and their descriptions, and model outputs without knowing which variant produced them. Each output was scored between 0 or 1 for factual accuracy, information missing from the codes, and relevance to the prediction task. CERE scored higher than both the model without structured conditioning and the model before RL refinement, providing evidence that it selects relevant information beyond the supplied codes.

Model variantFactual accuracyInformation beyond codesTask relevance
Without structured conditioning0.740.620.87
Before RL refinement0.800.710.88
CERE0.870.760.93

W2 Policy Refinement Stability and Sensitivity

Thank you for this comment. The concern regarding evidence quality is addressed in our response to W1. We have evaluated the model on MIMIC-III to examine sensitivity to the drift penalty and reference-policy penalty, which control representation changes and departures from the initial policy.

Penalty variedWeightAUROCPRAUC
Default settingsδ = 0.5, β = 0.0275.1338.43
Drift penaltyδ = 0.2575.0238.31
Drift penaltyδ = 1.074.9138.24
Reference-policy penaltyβ = 0.0174.9838.23
Reference-policy penaltyβ = 0.0474.9338.32
Rollout group sizePolicy updatesAUROCPRAUC
420074.8738.21
820075.1338.43
1620075.0838.39
810074.8138.17
840075.2438.59

As described in Appendix A.5, patients are sampled without replacement within an epoch, with one patient per update and eight rollouts per patient in the default setting.

Q1 Independent Evidence Quality Evaluation

Thank you for this question; we address the independent evaluation of evidence quality in our response to W1.

Q2 Policy Training Coverage, Sensitivity, and Reward Alignment

Thank you for these questions; we address these concerns in our responses to W1 and W2. Stage 3 policy refinement uses exactly 200 unique training patients per run, sampled without replacement from the training split. Each patient contributes eight rollouts, resulting in 1,600 patient-level rollouts across 200 optimizer updates. We will update the paper to reflect that.

Q3 Prediction Timing, Patient Selection, and Leakage Prevention

Thank you for this question. As described in Sections 2.1 and 3.2, we use an admission and the patient’s earlier history to predict readmission within 15 days or medications at the next admission. We include only codes and notes available at the prediction cutoff, and excluding records from the target admission. We exclude patients with only one admission because medication labels require a subsequent admission. This limits our evaluation to patients with repeated admissions and may affect the readmission rate. The table below reports cohort statistics and readmission rates for patients with exactly one historical visit before the target admission. We will update the paper accordingly.

StatisticMIMIC-IIIMIMIC-IV
# of patients (= # of instances)5,1607,132
Avg. unique diagnoses per patient11.139.02
Avg. unique procedures per patient4.381.79
Avg. unique drugs per patient35.8219.18
# of medication labels3,0723,406
Readmission rate per instance11.03%19.90%

Q4 Confidence Intervals, Thresholds, Baseline Tuning, and Sampling

Thank you for these questions. Sections 3.1-3.2 and Appendix A.6 provide some of these details. We report mean performance over six seeds, with 95% confidence intervals calculated. Medication F1 is reported at the best threshold from a grid of thresholds spaced by 0.1, with a separate threshold selected for each method. Baseline hyperparameters were selected using greedy search. For MIMIC-IV, we uniformly sample 20% of eligible patients after excluding overlap with MIMIC-III and before the train/test split. This subset remains fixed across seeds and methods. We will update the paper to clarify these procedures.

AuthorsOpenReview ↗

Follow-up on our responses

Dear Reviewer qUP8, As the discussion period is coming to an end, we would appreciate your feedback on whether our responses address your concerns. If any points remain unclear or need further attention, we would be happy to address them. Thank you for your time.

Thank you for the clarifications and additional experiments. Your responses have addressed my concerns, and I maintain my score.

Dear Reviewer qUP8, thank you for your positive assessment and for taking the time to review our responses. If there is anything further you would like us to address, we would be happy to do so.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.