acceptodds
Under review as a conference paper at ICLR 2027

From Probabilities to Prescriptions: A Unified Dependency-Aware Set Decoder for Medication and Herb Recommendation

Abstract

Western medication and traditional Chinese medicine (TCM) herb recommendations differ in clinical theories, vocabularies, and safety knowledge, yet share a common set-valued prediction problem: mapping a clinical presentation to a variable-size prescription. Candidate-wise probabilities alone do not generally determine the Jaccard-optimal prescription, because the same marginals can arise from different candidate co-occurrence patterns. We therefore formulate both tasks as post-hoc set decoding and propose Dependency-Aware Prescription Selection (DAPS), a backbone-agnostic decoder for frozen probabilistic recommenders. DAPS uses two cross-fitted stages: the first calibrates candidate scores and constructs a provisional prescription, while the second refines these scores using set-conditioned context. Crucially, it separates behavioral co-selection and exclusion learned from recorded prescriptions from externally listed conflicts such as drug–drug interactions or herb incompatibilities. The refined probabilities are converted into a variable-size prescription through Jaccard-oriented cardinality selection, followed by a validation-selected global external-conflict budget. Across two electronic health record (EHR) and two TCM corpora, DAPS improves Jaccard in all 36 matched backbone–dataset evaluations, with average gains of and percentage points on the EHR and TCM corpora, respectively. Analyses characterize the sources and robustness of these gains, while highlighting the distinction between observational dependencies and external conflict knowledge.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
medication recommendation, herb recommendation, causal modeling, identifiability, electronic health records, traditional Chinese medicine

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

This paper proposes Dependency-Aware Prescription Selection (DAPS), a backbone-agnostic decoder for frozen probabilistic recommenders, which uses two cross-fitted stages to calibrate candidate scores and construct a provisional prescription, and to refine these scores using set-conditioned context.

Critical strengths

The research problem addressed by DAPS is well-founded, its theoretical motivation is more robust than that of typical papers on modular combinations, and the experimental coverage is quite extensive.

In the introduction, the author explicitly points out that existing recommendation models for Western and traditional Chinese medicine primarily focus on improving the representations and prediction scores of candidate medications; however, these approaches ultimately require converting candidate probabilities into a prescription set of variable size.

Therefore, the core question is: How to convert candidate-wise probabilities into a better prescription set accounting for dependencies between candidates, the cardinality of the set, and external conflicts, without retraining the backbone model?

This framing is sound and offers a more specific focus than simply proposing a generic new medication recommendation framework.

Critical weaknesses

The paper has not fully demonstrated that the observed performance gains stem primarily from the highlighted "dependency-aware set decoding", rather than from other factors such as a more robust probability calibrator, additional supervisory models, or the optimization of the number of prescribed items.

All backbones are augmented with an additional, separately trained prediction refinement pipeline; therefore, authors must also address the following question: Do these performance gains stem from the dependencies, or from the re-learning and calibration of the original model's predictions?

Questions for authors

I am focused on two questions:

First, is the methodological innovation primarily a matter of complex system integration, or does it represent a truly unique and irreplaceable new decoding principle?

Second, are the experimental gains sufficient to validate the dependency-aware mechanism, as opposed to simply reflecting more robust supervised learning applied during post-processing? Existing experiments demonstrate that DAPS, as a complete system, works. However, authors have not yet sufficiently proven that dependency-aware set modeling is the primary reason why DAPS works.

3 comments
AuthorsOpenReview ↗

Responses to Reviewer yahQ (1/2)

We appreciate the time the reviewer spent on our paper and the thoughtful questions raised. We address each of them in turn below.


W: “The paper has not fully demonstrated that the observed performance gains stem primarily from the highlighted ‘dependency-aware set decoding’, rather than from other factors such as a more robust probability calibrator, additional supervisory models, or the optimization of the number of prescribed items. All backbones are augmented with an additional, separately trained prediction refinement pipeline; therefore, authors must also address the following question: Do these performance gains stem from the dependencies, or from the re-learning and calibration of the original model’s predictions?”

A1: Thank you for stating this so directly. Both: re-learning and calibration are the larger source overall, and set dependence is a measured additional source on TCM. We decomposed the gain on eight backbone-dataset pairs (two backbones per corpus), adding components in order, so each factor you name has its own row (values are changes in Jaccard points):

ComponentEHR (4 pairs)TCM (4 pairs)
Number of prescribed items (cardinality on backbone scores)+0.18+0.50
Step-1 calibrator, no expert bank+2.36+0.35
Expert bank (additional supervised models)+2.12+1.70
Step 2, set-conditioned refinement+0.64+2.18
of which tied to the case's own provisional set+0.08+0.45
Step 3, conflict budget−0.250.00
Total+5.06+4.72

EHR: calibration and the expert bank explain almost all of the +5.06. Dependence on the case's own set adds +0.08.

TCM: the set-conditioned Step 2 is the largest component (+2.18 of +4.72). Of this, +0.45 is tied to the case's own provisional set, measured by the shuffled-set control (each case receives another case's provisional set; models, capacity and training unchanged).

Q1: “First, is the methodological innovation primarily a matter of complex system integration, or does it represent a truly unique and irreplaceable new decoding principle?”

A2: Thank you for this question. Both, in measurable parts. The table below separates what a simple decoding rule achieves from what the full post-hoc stage achieves.

TestResult
Sizing rule applied to backbone scores, no other component+0.34 on average (−0.04 to +0.90)
Our sizing rule against a fit-tuned global threshold on the same refined scores+0.02 / +0.13 / +0.21 / +0.38 on the four corpora
Full DAPS against the best fit-tuned simple decoding rule, 36 pairshigher on 36/36, mean +5.25, min +1.34
Removing the expert bank from full DAPS0.39 lost on average
Removing the hand-built visit channels0.01 lost on average
Removing the two set-completion networks1.27 lost on average (0.20 to 3.51)
Removing the expert bank and Step 2 together3.20 lost of the 4.89 mean gain

No simple decoding rule reaches DAPS (36/36 pairs, mean +5.25). The post-hoc stage is therefore not replaceable by a rule.

The principle: the size should be chosen under the metric. Proposition 1(ii) shows that equal marginals do not determine the Jaccard-optimal set; empirically a tuned threshold on the same refined scores comes within 0.02 to 0.38 of our sizing rule.

What no rule replaces is the cross-fitted refinement. The expert bank and the set-completion networks are substitutes, and removing both loses 3.20 of the 4.89 mean gain.

AuthorsOpenReview ↗

Responses to Reviewer yahQ (2/2)

Q2: “Second, are the experimental gains sufficient to validate the dependency-aware mechanism, as opposed to simply reflecting more robust supervised learning applied during post-processing? Existing experiments demonstrate that DAPS, as a complete system, works. However, authors have not yet sufficiently proven that dependency-aware set modeling is the primary reason why DAPS works.”

A3: Thanks for separating the system from the mechanism. We measured the mechanism directly. The table below gives the gain tied to each case's own set, the headroom with an oracle set, and the residual dependence in the data.

MIMIC-IIIMIMIC-IVTCM-OnsiteTCM-Internet
Set-specific gain (shuffled-set control)+0.14+0.09+0.43+0.37
Oracle headroom (recorded set as context)+0.89+1.63+13.36+20.60
Candidate pairs with residual correlation above 0.2 after Step 10.06%0.13%2.71%1.74%

The mechanism's gain follows the conditional dependence in the data. It is clear on TCM and small on the EHR corpora, with supervised refinement as the larger source overall. On MIMIC the inputs (diagnoses, procedures, previous medications) already explain most co-prescription, so even an oracle provisional set would add only 0.9 to 1.6 points. On TCM the structure exists (oracle headroom of 13 to 21 points) and DAPS already captures part of it, which marks where further gains lie. Across the eight pairs the set-specific gain is +0.04 to +0.14 on MIMIC and +0.37 to +0.58 on TCM.


We appreciate these questions and welcome any follow-up during the discussion.

AuthorsOpenReview ↗

Follow-up to Reviewer yahQ

Dear Reviewer yahQ,

Thank you for your valuable feedback on our submission. With the rebuttal period nearing its end, we would be grateful to know whether our responses have addressed your concerns, and whether there is anything further we can clarify. We truly value your assessment and look forward to hearing from you.

Best regards,

The Authors

Reviewer_s2MAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper formulates Western medication recommendation and traditional Chinese medicine herb recommendation as a shared variable-size set-prediction problem and argues that candidate-wise marginal probabilities do not generally determine the Jaccard-optimal prescription when labels are dependent. It introduces Dependency-Aware Prescription Selection (DAPS), a post-hoc pipeline that augments frozen backbone scores with per-candidate gradient-boosted experts, two cross-fitted score-refinement stages, corpus-derived co-selection and exclusion features, neural set-completion models, validation-selected Jaccard-oriented cardinality rules, and a final pooled external-conflict adjustment. The theoretical section establishes marginal insufficiency through a counterexample, derives a top-k expected-Jaccard rule and an L1 plug-in regret bound under conditional independence, and gives a second-order decomposition showing where conditional covariances enter expected Jaccard. Experiments on MIMIC-III, MIMIC-IV, TCM-Onsite, and TCM-Internet report positive Jaccard gains for all 36 matched backbone–dataset pairs, accompanied by component ablations, sensitivity analyses, limited uncertainty estimates, and extensive implementation documentation

Critical strengths
  1. The theory appears internally consistent. Proposition 1(ii) constructs two prescription distributions with the same marginals but different unique Jaccard-optimal predictions, and the construction looks valid. The exchange argument behind the top-k result under conditional independence is clear, as is the plug-in regret analysis. I also appreciate that the authors do not overstate Proposition 2: they explicitly acknowledge that the remainder may be large and that the result does not provide a finite-sample estimator.

  2. The paper focuses on an important problem that is often left implicit. Many recommendation models produce candidate scores, but they do not fully explain how those scores should be turned into a final set or how its size should be chosen. Studying that last step—from dense scores to a variable-size prediction—is worthwhile even when the underlying encoder is not state of the art.

  3. The matched evaluation setup is much more convincing than comparing numbers copied from different papers. In the main experiments, DAPS and the competing plug-in methods operate on the same frozen probability outputs and the same cohorts. DAPS improves Jaccard for every evaluated EHR backbone on both MIMIC datasets, and it also improves results across the native and transferred TCM groups.

  4. The paper handles the clinical and ethical scope carefully. It makes clear that the model predicts recorded prescribing behavior, not treatment effectiveness, and that observational associations should not be interpreted as pharmacological relationships. The authors are equally direct about the conflict budget: it is a pooled constraint, not a patient-specific safety guarantee. They also responsibly discuss the missing patient and practitioner identifiers in the TCM data, along with the risk of learning institutional prescribing biases.

Critical weaknesses
  1. The paper does not clearly separate the gains from “set decoding” from those produced by the substantial supervised machinery added on top of the backbone. DAPS is doing much more than applying a decision rule to frozen probabilities. It trains a gradient-boosted classifier for every output label using the original presentation features, adds two meta-models with 200 trees each, and includes two neural set-completion models with width 512.

  2. Stronger simple post-hoc baselines are needed before the value of such a complex decoder is clear. Native TCM rankers are decoded using the rounded mean prescription size from the training set. DAPS, by contrast, uses fit-set labels for calibration and searches over several choices, including the cardinality source, blending with the backbone count, scaling, and an offset. Using a common fixed count is understandable because the native rankers do not provide their own sizing rule, but it is not a competitive validation-tuned baseline. There is also a more direct concern: after the full step-2 refinement, a tuned global threshold comes within about 0.4 Jaccard points of PB or soft-Jaccard sizing in the representative experiments. Ordinary score refinement followed by threshold tuning may therefore explain most of the improvement

  3. The results suggest that genuinely case-specific set dependence contributes less than the paper’s headline framing implies. In the shuffled-set control, the portion of the step-2 gain tied to each case’s own provisional set averages only +0.02+0.02 and +0.06+0.06 Jaccard on MIMIC-III and MIMIC-IV. The corresponding gains are larger on the two TCM datasets, at +0.58+0.58 and +0.35+0.35, but still account for only part of the overall improvement. Much of what remains appears to come from extra case-level information provided by the neural models trained on the training partition, rather than from the provisional-set dependency context itself.

  4. The final conflict-adjustment stage is not a standard per-case decoder. It works across an entire batch of predictions, controls a pooled pairwise conflict rate, and assigns swaps through a global priority queue. As a result, the prescription produced for one patient may change depending on which other patients are included in the evaluation or deployment batch. No individual prescription is guaranteed to satisfy a conflict-rate constraint

  5. Uncertainty estimates are missing for most of the headline results. Three-seed ablations and paired bootstrap intervals are reported for only four representative backbone–dataset pairs, even though the main claim covers 36 matched evaluations. The reported intervals are reassuring and exclude zero. Still, they do not show the uncertainty around every gain in the main tables or around the average improvements reported across backbone groups

  6. The integrity of the TCM splits remains a meaningful limitation, even though the authors cannot fully resolve it with the available data. Reliable patient and practitioner identifiers are missing from the public releases, so splitting at the case level cannot rule out repeated patients, repeated practitioners, or closely related prescriptions appearing across training, fit, and test sets. That overlap could especially benefit the expert bank, relation statistics, and set-completion models. The manuscript is transparent about the problem, but it weakens the generalization claims that can be drawn from the TCM results.

Questions for authors
  1. The final scores barely change across the original backbones. On the shared MIMIC-III cohort, four different models all end up between 56.88 and 56.94 Jaccard. The same pattern appears on TCM-Internet, where the eight native backbones are compressed into a narrow range of 32.47–32.59. That makes me wonder whether the shared expert and meta-model stack is doing most of the predictive work rather than simply decoding the frozen backbone. Why were no post-hoc ablations included for: (i) backbone scores with cardinality selection only, (ii) backbone scores with both calibrators but no auxiliary expert bank, (iii) the auxiliary expert bank without backbone scores, (iv) the expert bank with a simple threshold or count rule, (v) DAPS without the set-completion networks, and (vi) DAPS without the raw presentation-derived expert features? The current “no expert” experiment uses a materially different joint architecture, and the paper itself notes that it is not a component ablation of post-hoc DAPS.

  2. The TCM datasets do not include reliable patient or practitioner identifiers, so case-level splitting cannot ensure that the training, fit, and test populations are truly independent. Could the authors measure exact and near-duplicate prescriptions, repeated symptom–prescription templates, and performance after stricter deduplication? Temporal or institution-level splits would also help where possible. If those checks cannot be performed, the generalization claims should be narrowed accordingly.

  3. The TCM comparison does not give the baseline rankers and DAPS an equivalent decoding and tuning budget. Native rankers use a fixed prescription size based on the training-set mean, while DAPS uses fit-set labels to choose its cardinality policy. Please compare every backbone–dataset pair against validation-tuned global thresholds, tuned fixed-kk, standard calibration followed by thresholding, and calibration followed by expected-Jaccard sizing.

  4. Conflict adjustment is performed globally across a collection of prescriptions, not independently for each case. Swaps are assigned using a shared priority queue to control a pooled conflict rate, which means one patient’s output may change depending on the other cases in the batch. There is also no per-prescription safety guarantee. Could the authors test sensitivity to batch size and composition, report the proportion of prescriptions containing at least one listed conflict, and provide a case-local or streaming version of the procedure?

  5. DAPS adds a large amount of supervised modeling on top of the frozen backbone: per-candidate tree experts, two cross-fitted meta-models, presentation-derived features, and neural set-completion networks. It is not only a decoding rule over existing probabilities. Could the authors report backbone-only, expert-only, backbone-plus-sizing, no-expert, and no-set-completion variants to show whether DAPS genuinely complements the backbone or mostly replaces its predictions?

6 comments
AuthorsOpenReview ↗

Responses to Reviewer s2MA (1/5)

We are grateful to the reviewer for the careful reading and the detailed comments. Our responses to each weakness and question are given below.


W1: “The paper does not clearly separate the gains from ‘set decoding’ from those produced by the substantial supervised machinery added on top of the backbone. DAPS is doing much more than applying a decision rule to frozen probabilities. It trains a gradient-boosted classifier for every output label using the original presentation features, adds two meta-models with 200 trees each, and includes two neural set-completion models with width 512.”

A1: Thank you for raising this. We ran the six variants of your Q1 on eight backbone-dataset pairs (two backbones per corpus). The nested attribution is in the general response; the leave-one-out view, with everything else refitted (values in Jaccard points):

Removed from full DAPSMean Jaccard lostMIMICTCM
Expert bank0.390.57 to 0.750.02 to 0.25
Presentation-derived channels0.01−0.07 to 0.06−0.01 to 0.01
Two set-completion networks1.270.20 to 0.930.98 to 3.51
Expert bank and Step 2 together3.201.54 to 3.561.44 to 6.80

The supervised components are substitutes and carry about two thirds of the gain. Removing the expert bank alone costs little because the set-completion networks also see the presentation. With neither, cardinality selection plus the Step-1 calibrator keeps +1.69 of the +4.89 mean gain. DAPS is therefore best described as supervised post-hoc refinement followed by metric-aware set construction.

W2: “Stronger simple post-hoc baselines are needed before the value of such a complex decoder is clear. Native TCM rankers are decoded using the rounded mean prescription size from the training set. [...] There is also a more direct concern: after the full step-2 refinement, a tuned global threshold comes within about 0.4 Jaccard points of PB or soft-Jaccard sizing in the representative experiments. Ordinary score refinement followed by threshold tuning may therefore explain most of the improvement”

A2: Thanks for pointing this out. All 36 pairs now have five simple baselines tuned on the same fit split: global threshold (1,001 candidates), fixed k (1 to 45), Platt or isotonic calibration with a threshold, calibration with expected-Jaccard sizing, and calibration with DAPS's own 1,485-configuration cardinality search. The table below compares DAPS with the best of the five per pair.

CorpusPairsDAPS minus best tuned simple baseline, mean (min)DAPS higherDAPS minus expert bank alonePB sizing minus tuned threshold, same Step-2 scores
MIMIC-III5+6.36 (+3.02)5/5+1.19+0.02
MIMIC-IV5+7.54 (+3.97)5/5+1.50+0.13
TCM-Onsite13+5.04 (+2.13)13/13+7.73+0.21
TCM-Internet13+4.16 (+1.34)13/13+2.98+0.38

No tuned simple rule reaches DAPS (36/36 pairs, mean +5.25). This holds even when each baseline's best variant is picked on the test set.

The margin comes from the refinement. On the same refined scores a tuned threshold is within 0.02 to 0.38 of PB sizing, while PB sizing needs no threshold tuned per pair.

The fixed size did not disadvantage the native TCM rankers. A fit-tuned fixed k changes their Jaccard by +0.14 on average (+0.02 to +0.23 over 16 rows).

AuthorsOpenReview ↗

Responses to Reviewer s2MA (2/5)

W3: “The results suggest that genuinely case-specific set dependence contributes less than the paper’s headline framing implies. In the shuffled-set control, the portion of the step-2 gain tied to each case’s own provisional set averages only (+0.02) and (+0.06) Jaccard on MIMIC-III and MIMIC-IV. The corresponding gains are larger on the two TCM datasets, at (+0.58) and (+0.35), but still account for only part of the overall improvement. [...]”

A3: Thank you for reading that control so carefully. The new ablation separates the three quantities for two backbones per corpus (MedAlign / SafeDrug on MIMIC, SMGCN / PresRecST on TCM):

CorpusStep 2 totalLost without the set-completion networksTied to the case's own set
MIMIC-III+0.14 / +0.560.20 / 0.51+0.14 / +0.04
MIMIC-IV+0.79 / +1.070.74 / 0.93+0.09 / +0.06
TCM-Onsite+3.87 / +1.483.51 / 1.28+0.58 / +0.43
TCM-Internet+2.24 / +1.112.03 / 0.98+0.42 / +0.37

On TCM the set-completion networks carry the gain, mostly as input-driven predictors. +0.37 to +0.58 of their 0.98 to 3.51 is tied to the case's own provisional set. With the recorded set as context the headroom is +13.36 / +20.60 on TCM-Onsite / TCM-Internet against +0.89 / +1.63 on MIMIC, so the dependence exists on TCM and is where set-conditioned decoding has the most room to grow.

W4: “The final conflict-adjustment stage is not a standard per-case decoder. It works across an entire batch of predictions, controls a pooled pairwise conflict rate, and assigns swaps through a global priority queue. As a result, the prescription produced for one patient may change depending on which other patients are included in the evaluation or deployment batch. No individual prescription is guaranteed to satisfy a conflict-rate constraint”

A4: Thank you for raising this. We measured the batch dependence and added a case-local variant that removes it. The first table shows the sensitivity to batch size; the second compares the global queue with the case-local variant.

MIMIC-III, batch size1832128512full
Cases differing from the full-batch output59.8%32.1%18.8%10.5%3.8%0%
Jaccard55.7156.7256.8656.9156.9456.97
Global queue → case-localMIMIC-IIIMIMIC-IVTCM-OnsiteTCM-Internet
Jaccard56.97 → 56.4155.27 → 54.3351.95 → 51.9232.47 → 32.47
Pooled conflict rate (%)6.93 → 6.816.53 → 6.431.33 → 1.310.003 → 0.000
Cases within their own allowance53.7% → 100%50.7% → 100%54.5% → 100%99.7% → 100%

A per-case allowance bounds every prescription and removes the batch dependence. It is calibrated on the fit split, uses no cross-case queue, and meets the same pooled budget for 0.57 / 0.94 / 0.03 / 0.004 Jaccard. At batch size 32, 62.1% of MIMIC-III patients receive more than one distinct prescription across 20 random batch compositions under the global queue; under the case-local rule none do.

The submitted pooled numbers hold. The aggregate is stable under the global queue (Jaccard sd 0.033 across partitions at batch size 32).

W5: “Uncertainty estimates are missing for most of the headline results. Three-seed ablations and paired bootstrap intervals are reported for only four representative backbone-dataset pairs, even though the main claim covers 36 matched evaluations. The reported intervals are reassuring and exclude zero. Still, they do not show the uncertainty around every gain in the main tables or around the average improvements reported across backbone groups”

A5: Thank you for this point. Paired bootstrap intervals now cover all 36 pairs and the group averages, as summarized below.

GroupPairsMean ΔJaccard95% intervalPairs with interval above zero
MIMIC-III5+5.61[5.35, 5.87]5/5
MIMIC-IV5+5.94[5.81, 6.07]5/5
TCM-Onsite13+6.74[6.64, 6.84]13/13
TCM-Internet13+5.75[5.69, 5.81]13/13
EHR average10+5.77[5.63, 5.92]10/10
TCM average26+6.24[6.19, 6.30]26/26

All 36 intervals lie above zero. Both group averages match the abstract, and per-pair gains agree with the submitted tables to within 0.07.

AuthorsOpenReview ↗

Responses to Reviewer s2MA (3/5)

W6: “The integrity of the TCM splits remains a meaningful limitation, even though the authors cannot fully resolve it with the available data. Reliable patient and practitioner identifiers are missing from the public releases, so splitting at the case level cannot rule out repeated patients, repeated practitioners, or closely related prescriptions appearing across training, fit, and test sets. That overlap could especially benefit the expert bank, relation statistics, and set-completion models. [...]”

A6: Thank you for this concern. We measured the overlap between test and training cases and re-evaluated after removing it. The first table gives the share of test cases with a duplicate in training; the second gives the DAPS gain after deduplication.

CorpusExact (input, prescription) duplicatePrescription with Jaccard ≥ 0.8 to a training oneRecords in a repeated input signature
MIMIC-III0.00%7.0%0.10%
MIMIC-IV0.45%25.7%2.41%
TCM-Onsite2.61%65.4%29.6%
TCM-Internet0.07%49.5%4.10%
Mean ΔJaccard, DAPS minus backboneFull test setWithout exact duplicatesWithout any near-duplicate prescription
EHR, 10 pairs5.775.795.91
TCM, 26 pairs6.246.525.07
TCM, 24 pairs without VITA6.266.214.90

The gain survives strict deduplication: unchanged on EHR, four fifths kept on TCM. Exact duplicates are rare. Near-duplicate prescriptions are common on TCM, and after removing every one of them the mean gain over the 26 TCM pairs is still +5.07 (+4.90 without the two VITA rows). The TCM claims are stated for case-disjoint generalisation.

Q1: “The final scores barely change across the original backbones. On the shared MIMIC-III cohort, four different models all end up between 56.88 and 56.94 Jaccard. [...] Why were no post-hoc ablations included for: (i) backbone scores with cardinality selection only, (ii) backbone scores with both calibrators but no auxiliary expert bank, (iii) the auxiliary expert bank without backbone scores, (iv) the expert bank with a simple threshold or count rule, (v) DAPS without the set-completion networks, and (vi) DAPS without the raw presentation-derived expert features? [...]”

A7: Thank you for listing the variants so precisely. We ran all six on eight pairs; the table below shows two representative pairs.

Variant (Jaccard)MIMIC-III / MedAlignTCM-Onsite / SMGCN
Frozen backbone53.8944.42
(i) backbone + cardinality selection only53.8644.51
(ii) both calibrators, no expert bank56.2251.85
(iii) expert bank only, no backbone scores55.6443.47
(iv) expert bank + tuned threshold55.5443.26
(v) without the set-completion networks56.7748.37
(vi) without presentation-derived channels57.0451.88
Full DAPS56.9751.88

On MIMIC the compression across backbones comes from the expert bank. Variant (iii) uses no backbone scores and is therefore identical for every backbone: it reaches 55.64 on MIMIC-III and 53.76 on MIMIC-IV, which is 79% / 75% of the full gain averaged over the five backbones of each corpus (48% on TCM-Internet).

On TCM-Onsite it comes from the set-completion networks. The expert bank alone is below the backbone (−15%); variant (ii) reaches the full score without it, and variant (v) loses 3.51.

Reviewer_wD6fOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses a limitation of existing medication and herb recommendation methods: candidate-wise probabilities alone are insufficient for optimal prescription selection, as they do not fully capture dependencies among candidates or determine the appropriate prescription size. To address this issue, it proposes DAPS, a post-hoc set decoder that combines auxiliary tree experts, dependency-aware score refinement, Jaccard-oriented cardinality selection, and a global external-conflict budget. Experiments on four Western medication and TCM herb datasets demonstrate consistent Jaccard improvements.

Critical strengths
  1. The paper offers a useful perspective by explicitly separating candidate scoring from prescription-set selection. The marginal-insufficiency analysis provides a clear motivation for considering information beyond individual candidate probabilities.
  2. The experiments are extensive, covering Western EHR and TCM datasets, multiple backbones, ablations, uncertainty estimates, and computational costs.
Critical weaknesses
  1. Complicated model design with insufficient justification. The proposed DAPS combines auxiliary tree experts, manually constructed features, two calibrators, and two neural set-completion models, but the necessity of these components and their interactions is not sufficiently justified. The ablations do not isolate their individual contributions, particularly those of the tree experts and set-completion models. Furthermore, Table 10 shows that substantial step-2 improvements remain even when the provisional sets are randomly shuffled across patients, with particularly small set-dependent gains on MIMIC. It remains unclear how much of the improvement can be attributed to dependency-aware decoding rather than additional supervised predictors and feature engineering.
  2. Limited clinical relevance of Jaccard-oriented prescription selection. Maximizing Jaccard against recorded prescriptions does not necessarily produce clinically beneficial prescriptions. Although DAPS considers externally listed conflicts, its global conflict budget does not ensure patient-specific safety, and the method does not explicitly assess medication necessity, treatment coverage, or medication burden. Moreover, the proposed conflict-driven swaps may replace medications without ensuring that their therapeutic roles are preserved. These limitations leave the clinical significance of the resulting set decisions unclear.
Questions for authors
  1. Could the authors provide clinically reviewed examples assessing whether newly added medications are necessary and whether conflict-driven replacements preserve treatment coverage? It would also be useful to examine the trade-offs among prescription overlap, medication burden, and patient-level safety.
  2. How can DAPS be applied to an autoregressive recommender such as COGNet [1], which predicts the next medication conditioned on previously generated medications, rather than each medication’s probability of appearing in the final prescription?
  3. Table 16 shows that PB sizing performs similarly to simpler alternatives, including soft-Jaccard and a global threshold. What justifies its additional complexity?
  4. Could the authors provide a more detailed analysis of the global conflict adjustment, including its separate runtime and the sensitivity to candidate-pool size?

References
[1] Conditional Generation Net for Medication Recommendation, WWW, 2022.

5 comments
AuthorsOpenReview ↗

Responses to Reviewer wD6f (1/4)

We sincerely thank the reviewer for the time and effort devoted to reviewing our paper. Please find our point-by-point responses below.


W1: “Complicated model design with insufficient justification. The proposed DAPS combines auxiliary tree experts, manually constructed features, two calibrators, and two neural set-completion models, but the necessity of these components and their interactions is not sufficiently justified. The ablations do not isolate their individual contributions, particularly those of the tree experts and set-completion models. Furthermore, Table 10 shows that substantial step-2 improvements remain even when the provisional sets are randomly shuffled across patients, with particularly small set-dependent gains on MIMIC. It remains unclear how much of the improvement can be attributed to dependency-aware decoding rather than additional supervised predictors and feature engineering.”

A1: Thank you for raising this. Each component is now isolated on eight backbone-dataset pairs (two backbones per corpus): added in order, and removed from the full pipeline with everything else refitted. Values are changes in Jaccard points.

Added in orderEHR (4 pairs)TCM (4 pairs)
Cardinality selection on backbone scores+0.18+0.50
Step-1 calibrator, no tree experts+2.36+0.35
Tree experts+2.12+1.70
Step 2, set-conditioned refinement+0.64+2.18
of which tied to the case's own provisional set+0.08+0.45
Step 3, conflict budget−0.250.00
Total+5.06+4.72
Removed from full DAPSMean Jaccard lostMIMICTCM
Tree experts0.390.57 to 0.750.02 to 0.25
Manually constructed visit channels0.01−0.07 to 0.06−0.01 to 0.01
Two set-completion models1.270.20 to 0.930.98 to 3.51
Tree experts and Step 2 together3.201.54 to 3.561.44 to 6.80

The tree experts and the set-completion models are substitutes. Removing one costs little because both see the presentation; removing both loses 3.20 of the +4.89 mean gain. The manually constructed channels contribute 0.01, so the pipeline can be simplified by dropping them.

Set dependence is measurable on TCM (+0.45) and small on EHR (+0.08). This matches Table 10 of the submission (Table 11 in the revision). The backbone remains necessary: without backbone scores the tree experts fall below the backbone on TCM-Onsite, where the set-completion models carry the gain.

AuthorsOpenReview ↗

Responses to Reviewer wD6f (2/4)

W2: “Limited clinical relevance of Jaccard-oriented prescription selection. Maximizing Jaccard against recorded prescriptions does not necessarily produce clinically beneficial prescriptions. Although DAPS considers externally listed conflicts, its global conflict budget does not ensure patient-specific safety, and the method does not explicitly assess medication necessity, treatment coverage, or medication burden. Moreover, the proposed conflict-driven swaps may replace medications without ensuring that their therapeutic roles are preserved. These limitations leave the clinical significance of the resulting set decisions unclear.”

A2: Thank you for this thoughtful question. We measured the three properties you name (medication necessity, burden and therapeutic role) on the main backbone of each corpus (MedAlign on MIMIC, PresRecST on TCM); the results are below.

MIMIC-IIIMIMIC-IVTCM-OnsiteTCM-Internet
Mean size: backbone / DAPS / recorded20.22 / 20.76 / 19.6915.59 / 14.69 / 13.9415.00 / 16.08 / 14.7312.00 / 15.18 / 12.37
Precision: backbone → DAPS69.10 → 70.6263.64 → 68.6763.64 → 64.4644.42 → 43.45
Recall: backbone → DAPS71.89 → 75.4472.58 → 74.0364.60 → 69.4445.83 → 53.60
Added items present in the recorded prescription50.1%52.7%42.2%29.1%
Removed items absent from the recorded prescription72.2%76.2%76.3%80.8%
Cases with a conflict-driven swap39.6%39.8%2.0%0.1%

Necessity and burden: DAPS gains mainly by removing well. 72% to 81% of removed items are absent from the recorded prescription. Precision and recall rise together on MIMIC-III, MIMIC-IV and TCM-Onsite, so the overlap is not bought with longer prescriptions.

On therapeutic role, we added a swap rule that only allows a replacement from the same ATC class (our labels are ATC level 3 codes, so levels 1 and 2 are testable):

Step-3 swap ruleMIMIC-III JaccardPooled DDISwaps within ATC-1 / ATC-2MIMIC-IV JaccardPooled DDISwaps within ATC-1 / ATC-2
Submitted rule56.976.9313.4% / 1.3%55.276.5315.0% / 1.0%
Same ATC level 1 only56.716.93100% / 7.3%54.966.53100% / 5.5%
Same ATC level 2 only56.236.93100% / 100%54.406.53100% / 100%

Therapeutic role: same-class swaps meet the same conflict budget at a small cost. The cost is 0.27 to 0.31 Jaccard at level 1 and 0.74 to 0.87 at level 2, with every swap kept inside the class.

Patient-specific safety: a case-local stage bounds every prescription. It meets the same pooled budget for 0.57 / 0.94 / 0.03 / 0.004 Jaccard on the four corpora, so each patient's listed-conflict rate is bounded individually.

AuthorsOpenReview ↗

Responses to Reviewer wD6f (3/4)

Q1: “Could the authors provide clinically reviewed examples assessing whether newly added medications are necessary and whether conflict-driven replacements preserve treatment coverage? It would also be useful to examine the trade-offs among prescription overlap, medication burden, and patient-level safety.”

A3: Thank you for this request. As a first step toward clinician review, we provide concrete cases selected by a fixed rule: every MIMIC-III test visit (MedAlign backbone) with at most 7 diagnoses, at most 14 recorded classes, at least one addition and at least one conflict-driven swap. This gives 9 visits whose mean Jaccard gain (+0.031) matches the full test set (+0.034); we show the largest, median and smallest gain. Medications are ATC level 3 classes. The three cases are listed below.

Largest gainMedian gainSmallest gain
Diagnosesalcoholic cirrhosis, alcohol dependence and withdrawal, chronic airway obstruction, depressive disorder, inguinal herniaHIV disease, history of urinary calculiend stage renal disease, hypertensive chronic kidney disease, acute posthemorrhagic anemia, hypocalcemia, thrombocytopenia, excessive menstruation
Jaccard, backbone → DAPS0.333 → 0.4710.333 → 0.3890.273 → 0.188
Added, in the recorded prescriptionR03A adrenergic inhalants; C03D potassium-sparing agentsD04A antipruriticsnone
Added, not recordedA04A antiemetics; N07B drugs for addictive disordersA04A antiemeticsA12A calcium; A12B potassium; C02D arteriolar vasodilators; A02A antacids; A04A antiemetics
RemovedC07A beta blockers; J01M quinolones; J01D other beta-lactams (none recorded)A12A calcium (not recorded)none
Conflict-driven swap, submitted ruleN06A antidepressants → A04A antiemeticsN05B anxiolytics (recorded) → N07BN02A opioids → C02D
Swap under the same-ATC-1 ruleN06A → N02A opioids (recorded), Jaccard 0.562N05B → N07B, unchangedno swap executed

Recorded additions match coded diagnoses, and removals agree with the record. In the first case both recorded additions correspond to a coded diagnosis (inhaled bronchodilator with airway obstruction, potassium-sparing diuretic with cirrhosis), and every removed class is absent from the record. Across all 9 visits, 14 of 16 removed classes are absent from the recorded prescription.

The same-class rule preserves treatment coverage. The submitted rule swaps the antidepressant class for an antiemetic in a patient with a coded depressive disorder; the same-ATC-1 rule of A2 instead selects a recorded nervous-system class and raises the case to 0.562.

Trade-offs (numbers from A2). Overlap costs +0.75 to +2.81 items over the recorded prescription, class-preserving swaps 0.27 to 0.87 Jaccard, and a per-patient conflict bound 0.004 to 0.94.

Q2: “How can DAPS be applied to an autoregressive recommender such as COGNet [1], which predicts the next medication conditioned on previously generated medications, rather than each medication’s probability of appearing in the final prescription?”

A4: Thank you for this question. We ran it on a frozen MIMIC-III COGNet checkpoint. Each candidate's score is its maximum probability over the decoding trajectory, the reduction COGNet's own evaluator uses for ranking; DAPS then runs unchanged. The results are below.

MIMIC-III, COGNetJaccardF1#Med
COGNet beam output53.0768.4321.94
+ Step 156.6471.4921.00
+ full DAPS57.1271.8921.48

DAPS adds +4.05 Jaccard on COGNet (95% paired-bootstrap interval [3.70, 4.40]). DAPS needs only one score per candidate, so any autoregressive recommender that exposes its step probabilities can be plugged in the same way; Step 1 calibrates the reduced score into an inclusion probability.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all three reviewers for the careful reviews. All new numbers use the submission's protocol (same splits, same frozen backbone exports, one fixed seed); intervals are 95% paired bootstrap over cases, 10,000 replicates.

1. To all reviewers

  • The headline result holds under every new check (Reviewers s2MA W2/W5/W6, wD6f Q2). DAPS improves all 36 pairs, each with a 95% interval above zero (EHR +5.77 [5.63, 5.92], TCM +6.24 [6.19, 6.30]). It is above the best of five fit-tuned simple decoding rules on 36/36 pairs (mean margin +5.25), also when each rule's best variant is chosen on the test set. The gain remains after removing every near-duplicate prescription (EHR 5.77 → 5.91, TCM 6.24 → 5.07), and it extends to the autoregressive COGNet (53.07 → 57.12).

  • Where the gain comes from (Reviewers wD6f W1, s2MA W1/W3/Q1/Q5, yahQ W/Q1/Q2). We ran the six variants listed in s2MA Q1 on eight backbone-dataset pairs, two backbones per corpus. Adding components in order:

ComponentMean ΔJaccardRange
Cardinality selection on backbone scores+0.34−0.04 to +0.90
Step-1 calibrator, no expert bank+1.35+0.20 to +3.19
Expert bank+1.91+0.34 to +3.00
Step 2, set-conditioned refinement+1.41+0.14 to +3.87
of which tied to the case's own provisional set+0.27+0.04 to +0.58
Step 3, conflict budget−0.12−0.43 to +0.00

Supervised refinement is the larger source; set dependence adds to it on TCM. The set-specific part is +0.04 to +0.14 on MIMIC and +0.37 to +0.58 on TCM. On MIMIC the expert bank alone, without backbone scores, reaches 79% / 75% of the full gain; on TCM-Onsite it falls below the backbone and the set-completion networks carry the gain (1.28 to 3.51 Jaccard lost when they are removed).

  • A case-local conflict stage removes the batch dependence (Reviewers wD6f W2/Q4, s2MA W4/Q4). Under the pooled queue at batch size 32, 62.1% of MIMIC-III patients receive more than one distinct prescription across batch compositions. A case-local variant with a fit-calibrated per-case allowance bounds 100% of cases, does not depend on the batch, meets the same pooled budget, and costs 0.57 / 0.94 / 0.03 / 0.004 Jaccard on MIMIC-III / MIMIC-IV / TCM-Onsite / TCM-Internet.

2. To each reviewer

  • Reviewer wD6f, on clinical relevance. Three automated proxies (therapeutic-class preservation of swaps, prescription-size distributions, precision of additions and removals), a swap rule restricted to the same ATC class, three worked cases chosen by a fixed rule, and a measured result on COGNet.
  • Reviewer s2MA, on simple baselines, uncertainty and TCM split integrity. Five tuned simple baselines on all 36 pairs, intervals for all 36 pairs and both group averages, duplicate and near-duplicate counts with results after strict deduplication, and a temporal split on TCM-Onsite with a size-matched control.
  • Reviewer yahQ, on principle versus system integration. The attribution above split by medical system, the leave-one-out view, and the shuffled-set control.

3. General changes in the revised paper

All changes are marked in blue.

  • Abstract, introduction and conclusion state the attribution above: cross-fitted supervised refinement is the larger source of the gain, and set-conditioned dependence is a measured contribution on TCM.
  • Section 3, Step 3 adds the case-local conflict variant, which bounds every prescription individually, and the same-ATC-class swap rule.
  • Section 4 has a new paragraph, "Attribution and Robustness of the Gains", with the component attribution in a new Table 5, which complements the step-wise ablation of Table 4. It summarizes the tuned simple baselines, the intervals for all 36 pairs, the deduplicated results and COGNet.
  • Table 3 and Section 4 note the unified input protocol for the TCM rows. The rows for Western architectures transferred to TCM used an export that also carries later-visit text; under the first-visit input of the native rows (new Table 46), the eight rerun pairs reach the native group's level (51.59 against 51.57 on TCM-Onsite; 32.72 against 32.50 on TCM-Internet), DAPS improves each of them by +10.58 / +10.08 on average, and all 36 pairs remain positive.
  • A new Appendix F, "Additional Experiments and Analyses", with 24 tables: F.1 component attribution; F.2 tuned simple baselines on all 36 pairs; F.3 intervals for all 36 pairs (Table 34); F.4 conflict stage (batch sensitivity, case-local variant, runtime, class-preserving swaps); F.5 clinical proxies and worked cases; F.6 COGNet; F.7 duplicates and the temporal split; F.8 unified input protocol.
  • Appendix A.4 (Limitations) adds the per-prescription safeguards, the need to refit DAPS together with the backbone under temporal drift, and the case-disjoint scope of the TCM results.

We are glad to clarify anything further or to run additional experiments during the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.