acceptodds
Under review as a conference paper at ICLR 2027

CoREL: Conditional Relative Evidence Learning for Multi-Institutional Medication Recommendation

Abstract

Medication recommendation in real-world clinical settings must account for heterogeneity in prescribing practices across healthcare institutions. Existing methods typically fit observed prescriptions directly, potentially conflating patient-dependent prescribing evidence with institution-specific prescribing tendencies. We propose CoREL, a multi-institutional medication recommendation framework that separates shared evidence learning from institution-specific prescription calibration. Under an additive decomposition of prescription decisions, we introduce conditional relative evidence learning to eliminate institution–medication tendencies while enabling shared patient-dependent prescribing evidence to be learned across institutions. Dual-level calibration then establishes a global prescribing reference and estimates institution-specific tendencies, recovering institution-conditioned prescription probabilities. An inference-time refinement procedure further balances prescription support against known harmful drug–drug interactions (DDIs). Experiments on eICU, MIMIC-IV, and a real-world two-center inpatient dataset demonstrate improved recommendation accuracy and lower DDI rates, with particularly pronounced gains in multi-institutional settings. Code is available at https://anonymous.4open.science/r/CoREL-01C5.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Medication recommendation, Electronic health records, 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_kcKiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CoREL addresses medication recommendation when hospitals differ in prescribing practices. It assumes that each medication's prescription logit is the sum of a shared patient-dependent score and a hospital-specific intercept. Training compares patients from the same hospital with discordant medication labels, conditioning out that intercept. A GRU-based EHR encoder learns the shared scores. Then subsequent global and shrinkage-based local intercept calibration recovers hospital-conditioned predictions. Greedy inference-time refinement discourages known drug-drug interactions. Experiments shows CoREL generally improves recommendation accuracy and reduces DDI rates in multi-institutional settings.

Critical strengths
  1. Practical motivation: Figure 1 shows substantial hospital variation in prescribing therapeutically related medications. CoREL addresses this by removing a medication-specific intercept for within-hospital comparisons, while retaining useful local practice information in later calibration
  2. The paper provides enough details for reproducibility.
Critical weaknesses
  1. Limited technical novelty and missing relevant baselines. CoREL combines a shared scorer and hospital-specific intercepts (Eq 1), a standard pairwise logistic loss (Eq 5), and intercept calibration with shrinkage toward a global reference. The baselines also omit TEMPT [1], which directly addresses multi-center medication recommendation on eICU through shared pretraining and hospital-specific prompt tuning. [1] A Contrastive Pretrain Model with Prompt Tuning for Multi-center Medication Recommendation. Liu et al, TOIS 2025
  2. Insufficient validation of the claimed separation. The paper lacks controlled evidence that the learned scores separate patient evidence from institutional tendencies beyond the additive-intercept assumption in Eq 1. In particular, recovery of the two components and robustness to institution-by-patient interactions remain untested. Prescription-based prediction metrics cannot independently establish this separation. Appendix B.1 also leaves the prediction-time feature cutoff unclear, so whether current-encounter diagnoses and procedures precede the target medications requires clarification.
Questions for authors
  1. What is the main technical advance over conditional pairwise learning followed by regularized intercept calibration?
  2. How does CoREL compare with TEMPT under matched inputs and data splits?
  3. Can controlled experiments assess component recovery and robustness to institution-by-patient interactions?
  4. When is the recommendation made, and how are current-encounter inputs restricted to information available at that time?
Reviewer_35oPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes CoREL, a multi-institutional medication recommendation framework that separates shared patient-dependent prescribing evidence from institution-specific prescribing tendencies. It introduces conditional relative evidence learning using within-institution patient comparisons to reduce the influence of institutional biases during training. A global-to-local calibration strategy incorporates institution-specific prescribing tendencies, followed by DDI-aware refinement to generate the final medication set. Experiments on multiple real-world EHR datasets demonstrate improved recommendation accuracy and reduced DDI rates.

Critical strengths
  • The paper addresses an important problem of institutional heterogeneity in multi-institutional medication recommendation.
  • The proposed framework is mathematically well-motivated under its additive decomposition assumption.
  • Extensive experiments across multiple datasets demonstrate consistent improvements in recommendation accuracy and DDI rates.
  • The ablation studies and cross-institutional generalization experiments provide empirical support for the proposed components.
Critical weaknesses
  • The core approach primarily builds on established conditional likelihood, pairwise ranking, and calibration techniques, without introducing substantial methodological advances.
  • Although institution-specific calibration improves agreement with historical prescriptions, it remains unclear whether incorporating institutional prescribing tendencies leads to clinically more appropriate recommendations.
  • The framework assumes that institutional prescribing tendencies are patient-independent additive biases, overlooking potential interactions between institutional policies and patient characteristics.
  • The experiments do not directly establish whether patient-dependent prescribing evidence is meaningfully separated from institution-specific preferences.
  • Improvements in prescription matching and DDI rates do not necessarily demonstrate improved treatment effectiveness or patient-level clinical safety.
Questions for authors
  • While the experiments demonstrate that institution-specific calibration improves agreement with observed prescriptions, the clinical significance of this improvement remains unclear. If clinically comparable patients receive different medication recommendations solely because they belong to different institutions, how can the authors establish that these differences reflect clinically justified decisions rather than institutional prescribing habits or non-clinical preferences? Could the authors provide an evaluation based on independent clinical appropriateness criteria, rather than historical prescription agreement alone?
  • Equation (1) assumes that institution-specific prescribing tendencies can be represented by a patient-independent additive term. However, institutional policies may interact with patient characteristics, such as disease severity, comorbidities, and demographics. Under such conditions, the proposed within-institution comparisons may not fully eliminate institutional effects. Could the authors evaluate the robustness of CoREL under patient-dependent institutional shifts, either through controlled synthetic experiments or subgroup-specific analyses of real-world prescribing patterns?
  • Figures 3 and 4 demonstrate improved cross-institutional ranking performance and stable institution-specific calibration, but these results do not directly establish that patient-dependent evidence and institutional tendencies are effectively disentangled. Could the authors conduct controlled experiments with known institutional biases to quantify how accurately the proposed method separates the two components? Additionally, how sensitive is the learned shared evidence to differences in patient populations across institutions?
  • The conditional relative learning objective in Equations (3)–(6) closely follows established conditional likelihood and pairwise logistic ranking principles. Although Table 3 compares CoREL with institution-specific BCE learning, it does not clearly establish the benefits of the proposed objective over conventional conditional logistic or pairwise ranking approaches under the same calibration procedure. Could the authors provide such comparisons and clarify what methodological advance is introduced beyond adapting these existing principles to medication recommendation?
  • The learned institution-specific terms may reflect not only clinically justified preferences but also drug availability, formulary restrictions, procurement policies, or historical prescribing habits. Could the authors analyze whether the estimated institutional tendencies correspond to clinically meaningful prescribing differences? In particular, would changing the institution identifier while keeping the patient state fixed lead to different medication recommendations, and if so, what proportion of these changes represent clinically interchangeable substitutions versus potentially inappropriate treatment differences?
  • The proposed greedy selection procedure does not generally guarantee a globally optimal medication set under the implicit objective combining prescription scores and pairwise DDI penalties. Could the authors compare the proposed greedy strategy with alternative optimization methods, such as integer programming or beam search, under the same objective? In particular, it would be useful to evaluate whether the greedy approximation introduces meaningful losses in recommendation accuracy or the F1–DDI trade-off, and whether its computational efficiency justifies these potential limitations.
Reviewer_jm8VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a multicenter drug recommendation model, CoREL, which splits prescription log-probability to eliminate institutional bias and provides recommendations through greedy optimization based on calibration and drug interactions. The model was tested in eICU, MIMIC-IV, and a self-built cohort, achieving performance improvements over REFINE. Cross-institutional generalization and ablation experiments were also conducted. The research topic is reasonable, statistically consistent, and the multicenter evaluation is a highlight; however, significant questions remain regarding temporal leakage, the definition of institutional effects, and gain attribution.

Critical strengths

The characterization of multicenter prescription heterogeneity is intuitive and interpretable. Cross-hospital prescription aggregation can confuse patient conditions with local prescribing habits. This paper explicitly separates these two aspects using additive log-odds and site biases from inconsistent prescriptions. While the conditional likelihood method itself is not novel, its combination with global-local calibration to form a complete cross-site recommendation scheme is the paper's greatest strength.

The multicenter assessment is more comprehensive than similar studies. The model shows a considerable improvement over REFINE in eICU and its transferability is validated in 26 hospitals that did not participate in the training. Small-sample calibration shows good results. Additional real-world two-center cohorts, ablation experiments, and time consumption tests were also included. However, some core conclusions still require further validation before they can be accepted as reliable.

Critical weaknesses

The shared effect hypothesis and temporal validity lack sufficient validation. The model assumes that institutional differences are only reflected as additive drug biases and that patient characteristic effects remain unchanged across sites, an assumption that is relatively strong. In-hospital prescription standards and case composition can introduce interaction effects; paired targets only eliminate biases and cannot guarantee that learned clinical indicators remain unchanged across sites. Furthermore, the lack of a clear prediction cutoff point leads to temporal leakage: some diagnoses and procedures are entered only after medication is prescribed. Retrospective modeling is not equivalent to prospective clinical decision support; feature timestamps need to be validated and key experiments redone based on valid temporal inputs.

Performance gains are difficult to attribute to CREL, and several analyses require revision. Existing ablation controls may not be fair, and variables such as encoders and calibration procedures are not aligned. The correlation test in Figure 4(b) may use the same fitted data, lacking out-of-sample calibration evidence. The AP index for drug interaction ablation is contradictory: if DDI optimization only modifies the final drug selection without changing the probability ranking, AP should not decrease. A matched baseline must be established, out-of-sample results supplemented, and anomalies explained to confirm that the performance improvement originates from CREL.

Questions for authors

Q1: Explain the timestamp rules and temporal constraints used, list the source tables and timestamps corresponding to diagnoses, operations, and features in each dataset, and explain the handling methods for post-discharge diagnoses and post-prescription operations; clarify whether the task is retrospective reconstruction or prospective recommendation at the time of prescription.

Q2: Remove the gain from CREL and verify the institutional additive effect hypothesis. Confirm that BCE ablation with institutional logit maintains consistency in model capacity, calibration, contraction strategy, and DDI inference; supplement the results of shared encoder plus institutional bias, quantify the cross-hospital differences or subgroup calibration of patient feature effects, and explain whether Figure 4(b) uses the training set or reserved data.

Q3: Explain the AP contradiction in DDI ablation and explain the implementation of the balanced version of CREL. If DDI optimization only applies to post-threshold screening, explain the reasons for AP changes; describe the intra-hospital small-batch weighting scheme to satisfy the hierarchical balance target of Equation (6), especially for rare drugs with inconsistent pairings.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.