Pith. sign in

REVIEW 3 major objections 6 minor 59 references

This paper claims that LLM-generated counterfactual patient records, instruction-tuned clinical summarization, and token-by-token medication decoding lift rare-medication recommendation by up to 30.9% over 14 baselines.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

GenRxR improves rare-medication recommendation using LLM-generated counterfactual data, instruction tuning, and sequential medication generation.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Solid and genuinely new empirical recipe for rare-med recommendation, but the output-length confound and thin synthetic-fidelity check keep it from being fully settled. the 3 major comments →

arxiv 2607.24829 v1 pith:VS7RZYJ5 submitted 2026-07-23 cs.IR cs.LG

Improving Rare Medication Recommendation with Counterfactual Data Augmentation and Large Language Models

Shinhwan Kang , Soo Yong Lee , Jaewon Kim , Kijung Shin , Buru Chang This is my paper
classification cs.IR cs.LG
keywords Rare medication recommendationcounterfactual data augmentationlarge language modelselectronic health recordsinstruction tuninglong-tail distributionMIMIC-IIIMIMIC-IV
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Existing AI medication recommenders predict commonly prescribed drugs well but frequently miss rarely prescribed ones. The paper attributes this to two shortcomings: a scarcity of rare-med training instances and a lack of attention to how medications are co-prescribed together. GenRxR addresses both with a two-LLM pipeline: a counterfactual generator (GPT-4o) invents plausible medical records in which a patient is assumed to have received a rare drug, and a recommendation LLM (LLaMA 3.1 8B) is first instruction-tuned on clinical summarization and then fine-tuned to output medications sequentially, conditioning each prediction on the drugs already recommended. On MIMIC-III and MIMIC-IV, this framework beats 14 baselines, improving rare-med predictive performance by up to 30.9% (Group 1 Jaccard from 10.19 to 13.27 on MIMIC-III) while keeping frequent-med and micro-averaged performance essentially intact.

Core claim

The paper's central claim is that LLM-generated counterfactual data plus instruction tuning for clinical summarization plus autoregressive medication prediction is a workable answer to the long-tail medication problem. The quantitative evidence is Table 1: GenRxR outperforms all 14 baselines on macro-averaged Jaccard and F1 on both datasets, with the largest relative gains concentrated in the rarest medication groups (up to 30.9% over COGNet on MIMIC-III Group 1). The ablation shows both counterfactual data and instruction tuning contribute, individually raising Group-1 Jaccard from 10.21 to 12.29 and 11.98, respectively, on MIMIC-III. A plausibility test in which a human expert (accuracy 0.

What carries the argument

The load-bearing mechanism is a two-LLM pipeline. The counterfactual generator LLM_CF selects, for each target rare medication m*, a set C_m* of associated diagnoses, procedures, and co-medications ranked by relative risk (P(c|m*)/P(c|¬m*)), constructs a hypothetical 'what-if' prescription for patients who never got m*, and prompts GPT-4o to pick the most plausible of three candidate patients and modify that patient's record for medical consistency. The instruction-tuned model LLM_IT (LLaMA 3.1 8B, LoRA) is trained to summarize medical records with respect to a target medication, using LLM_CF's summaries as ground truth. The recommendation model LLM_REC (same base, further fine-tuned) output

Load-bearing premise

The entire rare-med gain rests on the assumption that GPT-4o's counterfactual records—biased by relative-risk-selected clinical element sets—are clinically plausible and free of label leakage, meaning they preserve the true conditional distribution of diagnoses, procedures, and medications given patient history; the paper itself flags relative risk as unreliable for very rarely prescribed medications and defers larger-scale clinical validation to future work.

What would settle it

Train a classifier to predict the target medication m* from the other clinical elements (diagnoses, procedures, co-medications) of generated counterfactual records. If the classifier's AUC on held-out generated records is well above chance (e.g., >0.8), the counterfactual generator encodes the label into the features, and the downstream rare-med improvement is a shortcut rather than genuine clinical reasoning; the paper does not report such a leakage check.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Rare-med groups gain roughly 23–31% in Jaccard/F1 on MIMIC-III/IV, while micro-averaged metrics stay within about 1.8% of the strongest baseline, so the improvement is targeted rather than a uniform lift.
  • Ablation shows counterfactual data alone raises Group-1 Jaccard from 10.21 to 12.29 on MIMIC-III, and instruction tuning alone raises it to 11.98; combining both gives 13.27.
  • The predicted sets have DDI ratios (8.29 MIMIC-III, 7.71 MIMIC-IV) close to ground truth (8.26, 7.89), suggesting rare-med gains do not come with markedly worse drug-drug interaction safety.
  • A blind test with a human expert (accuracy 0.45) and LLM evaluators (0.50–0.61) finds generated counterfactual records hard to distinguish from original records, supporting clinical plausibility.
  • The sequential (token-by-token) decoding, with ground-truth medication order sorted by ascending frequency, is a simple mechanism that may be reused in any generative medication recommender.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension would be to generate counterfactual data with an open-source LLM instead of GPT-4o; if the rare-med gains persist, the approach becomes reproducible in privacy-constrained hospital settings where proprietary APIs may be unavailable.
  • The relative-risk element selection is computed only from training admissions, so the augmentation cannot create associations absent from the data; a distributional check of P(diagnoses, procedures, medications) in generated vs original records would tell whether the LLM amplifies existing correlations or discovers new clinically meaningful patterns.
  • The same recipe—LLM counterfactuals, summarization-based instruction tuning, autoregressive decoding—could transfer to other sparse-label healthcare tasks such as rare diagnosis prediction or rare procedure recommendation; the paper does not test these transfers.
  • A latent risk is label leakage: if the counterfactual generator encodes m* into the modified diagnoses/procedures, a tuned reader could recover the target medication from the features, making the recommendation gain a shortcut; the paper reports no leakage test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes GenRxR, an LLM-based framework for rare medication recommendation. It generates counterfactual EHR records with GPT-4o using relative-risk-derived clinical element sets and patient candidates, uses these records and LLM-generated summaries for instruction tuning of LLaMA-3.1-8B, and then fine-tunes the model to generate medications sequentially. On MIMIC-III and MIMIC-IV, it reports large relative gains for the rarest medication group (up to about 30.9% in Jaccard/F1) over 14 baselines while keeping micro-averaged performance comparable. The paper includes ablations, DDI ratios, a plausibility discrimination test, and qualitative examples.

Significance. If the effect is real and not an artifact of decoding length or sequence ordering, this is a meaningful contribution: it demonstrates a practical way to use LLM counterfactual generation to mitigate long-tail scarcity in a high-stakes recommendation task, and it provides the first counterfactual augmentation pipeline for medication recommendation. The empirical apparatus is above average: two public EHR datasets, 14 baselines, three random trials with standard deviations, ablations separating CF and IT, DDI safety reporting, human/LLM plausibility evaluation, and a promised code/data release. The main results, however, are not yet conclusively separated from output-count and sequence-ordering effects, and the plausibility evidence for synthetic data is indirect. These are addressable with additional experiments, so the paper merits major revision rather than rejection.

major comments (3)
  1. [§5.2, Table 1 and Table 2] The headline Group 1 improvements are not controlled for the number of predicted medications. On MIMIC-IV, GenRxR outputs 13.87 medications/admission vs 11.58 for FLAME (the strongest Group 1 baseline) and 11.46 in the ground truth; on MIMIC-III it outputs 23.22 vs 19.58 ground truth, and in the ablation (Table 2) adding CF raises Group 1 Jaccard from 10.21 to 12.29 while #Predicted rises from 21.95 to 23.19. Jaccard/F1 are count-sensitive, so the +22.8–30.9% rare-med gains could be partly a more permissive output length. Since §4.3 gives no length cap or decoding threshold, and the paper reports no precision-recall curves or matched-count evaluation, the central claim is not yet established. Please add per-group precision/recall and a matched-count control where GenRxR is constrained to each baseline's output cardinality.
  2. [§4.3 (footnote 3)] The training objective converts each medication set into a sequence ordered by ascending frequency in the training set, explicitly to prioritize rarely prescribed medications. A sequential generator trained this way will learn to emit rare meds early; at inference the model is not required to justify the inclusion of a rare med by patient evidence. This ordering, rather than CF data or co-medication modeling, may explain part of the rare-med gain. No ablation swaps the ordering. Please report results with frequency-descending or random medication order, and/or show that rare-med improvements persist when the model is forced to allocate a fixed output budget.
  3. [§4.1, §5.4 and §6] The generated counterfactual records are the core of the method, yet their validation is only a discrimination test. Table 3's accuracy near chance shows the records are statistically hard to distinguish, not that they are clinically correct or distributionally faithful. The LLM evaluators are from the same vendor family as the generator; the human evaluator is a single expert on 100 samples. Footnote 2 concedes that relative risk may be unreliable for very rare meds, and §6 concedes that drug-disease conflicts are not screened. Given that CF data contributes the largest ablation gain (Table 2), the paper should quantify the risk of hallucinated or unsafe additions, e.g., by checking whether generated diagnoses/procedures are consistent with the target medication's known contraindications, and by measuring downstream predictive performance when the synthetic data are filtered by a clinic
minor comments (6)
  1. [§2 heading] Typo: 'Medication Recommandation' should be 'Medication Recommendation'.
  2. [General] The 'online appendix' is referenced many times (e.g., §5.1, §5.4, §5.5) but reference [24] is only a GitHub URL; no appendix document is included in the arXiv submission. Please provide the appendix in the same package.
  3. [§5.1] The rare-med group definitions are arbitrary: 28 per group in MIMIC-III and 30/30/30/32 in MIMIC-IV. Consider additionally reporting per-medication results for the rarest 1–5% of medications or specifying a frequency threshold with clinical rationale.
  4. [Table 1] The 'GT' entries under DDI Ratio (8.26 and 7.89) are easy to confuse with model predictions; consider moving ground-truth values to a separate row or a table note.
  5. [§5.1] Hyperparameters such as k=3, top-15 diagnoses/procedures, top-10 medications, and temperature 0.3 are reported, but no sensitivity analysis is given for the size of C_m*. The appendix reportedly covers LoRA-rank sensitivity only; a brief sensitivity table for these choices would strengthen the paper.
  6. [§5.2] Standard deviations are reported across three trials, but no statistical significance tests are provided. Pairwise tests (e.g., corrected paired t-tests) would clarify which differences are reliable, especially for the smaller Group 1 gains on MIMIC-IV.

Circularity Check

0 steps flagged

No circular derivation: the reported rare-med gains are evaluated on held-out real MIMIC records, not on the counterfactual training data.

full rationale

The paper's chain is: select associated clinical elements C_m* by relative risk from the training split (Section 4.1); prompt GPT-4o to generate counterfactual admissions that include the assumed m*; instruction-tune LLaMA on summaries and generated data (Section 4.2); fine-tune the recommendation model on original plus augmented EHR data (Section 4.3); then evaluate Jaccard/F1 on held-out MIMIC-III/IV admissions (Sections 5.1-5.2). No equation or construction step in that chain substitutes its conclusion into its premises. The counterfactual augmentation intentionally re-encodes training co-occurrence statistics into the training set, but the headline 30.9% rare-med gain is measured against real held-out admissions and 14 external baselines, so it is an empirical finding rather than a constructional identity. The paper's own footnote that relative risk may be unreliable for very rare medications is a data-quality caveat, not a circular reduction. The only author self-citation, [7], supports threshold conversion in the plausibility probe and is not load-bearing. The skeptic's output-length / #predicted concern is a possible confound or correctness issue (unmatched prediction counts, no precision-recall or length-sensitivity analysis), not a circular step, because it does not show that the test-set predictions equal the training inputs by construction.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The framework introduces no new theoretical entities, but it does introduce a new data-generation pipeline whose validity is carried by (a) a hand-chosen set of relative-risk context elements, (b) an unverified assumption that GPT-4o produces clinically coherent counterfactual records, and (c) several hand-set hyperparameters (k=3, 30 scenarios/medication, temperature=0.3, top-15/15/10 context). Five free parameters and two invented data artifacts mean the reader must accept a considerable amount of postulation beyond the cited baselines. The central claim is not definitionally circular, but the absence of any external validation of the synthetic data is the largest burden.

free parameters (5)
  • Clinical element context size (top-15 diagnoses, top-15 procedures, top-10 medications per target medication) = 15 / 15 / 10
    Chosen by hand in Section 5.1 ('for each generation, we selected 15 associated diagnoses, 15 procedures, and 10 medications in descending order of relative risk'). Not swept in the main paper; this controls how much label-side information the counterfactual generator sees.
  • Number of candidate patients k = 3
    Section 5.1: 'involves randomly sampling three patient candidates... (i.e., k=3)'. Chosen by hand; affects the diversity and plausibility of counterfactual generation.
  • Number of counterfactual scenarios per medication = 30
    Section 5.1: 'we generated 30 counterfactual scenarios for each medication'. Chosen by hand; has a direct effect on the augmentation ratio and the rare-group gains in Table 1.
  • Temperature of GPT-4o during counterfactual generation = 0.3
    Section 5.1: 'utilized GPT-4o for counterfactual inference model LLM_CF with a temperature of 0.3'. Affects the diversity/plausibility trade-off of generated records.
  • LoRA rank / alpha for instruction tuning and recommendation fine-tuning = not stated in main text
    Section 5.5 refers to a 'LoRA-rank sensitivity analysis' in the online appendix only. In the main paper, the LoRA rank is an unstated hyperparameter that could affect the magnitude of all reported gains.
axioms (5)
  • domain assumption GPT-4o's counterfactual outputs preserve the joint distribution of diagnoses, procedures, and medications in a clinically coherent way.
    Invoked throughout Section 4.1 and validated only by a 100-sample blind discrimination test (Section 5.4) with a single human expert (accuracy 0.45). If false, the augmented training set could inject implausible patterns that inflate rare-med metrics.
  • domain assumption MIMIC-III/MIMIC-IV pre-processing and 4:1:1 split following GAMENet/COGNet/SafeDrug protocols is exactly comparable across all baselines.
    Section 5.1: 'we followed the same data pre-processing and dataset split protocol as in previous studies'. The paper does not re-run baselines under its own codebase, so cross-paper numbers inherit subtle differences in medication coding, visit truncation, and split seeds.
  • domain assumption The relative-risk-based C_m* sets capture clinically sufficient conditions for recommending m*.
    Section 4.1: 'A high relative risk indicates that a clinical element frequently co-occurs with the target medication m* but is uncommon otherwise, making it a strong candidate for inclusion in C_m*.'. The paper itself notes relative risk 'may be unreliable for medications observed in very few samples', which is exactly the rare-meds regime.
  • domain assumption LLaMA 3.1 8B Instruct, after LoRA instruction tuning and LoRA fine-tuning, can learn token-level sequential medication generation from MIMIC data.
    Section 4.3 defines the objective as next-token prediction over medication sequences, with medications ordered by ascending frequency. This assumes that the LLM's text prior does not dominate the ID+text embeddings for rare codes, and that greedy/beam decoding of the medication sequence recovers a set that matches ground-truth sets.
  • domain assumption Jaccard and F1, as computed over medication sets from generated text, are not biased by the model's tendency to predict more medications.
    Table 1 shows GenRxR predicts 23.22 medications per admission on MIMIC-III vs ground-truth 19.58 (+18.6%) and 13.87 vs 11.46 on MIMIC-IV (+21%). Jaccard penalizes false positives, but a systematic over-prediction bias still inflates recall-driven F1 for rare groups unless the precision loss is fully captured; the paper does not analyze this trade-off per group.
invented entities (2)
  • Counterfactual clinical records R-hat* generated by LLM_CF no independent evidence
    purpose: Synthetic training examples that assume a target medication was prescribed to a patient who did not receive it, with LLM-modified diagnoses, procedures, and co-medications.
    These records are new data objects introduced by the paper. The only independent evidence of their validity is the Section 5.4 blind test (single human expert, accuracy 0.45; LLM judges 0.50–0.61) and the qualitative Table 4 analysis. There is no predictive or external clinical validation.
  • Medical record summaries O_IT generated by LLM_CF and used as ground truth for instruction tuning no independent evidence
    purpose: Training targets for LLM_IT's summarization task, intended to align the model with medication-relevant clinical reasoning.
    The summaries are generated by the same proprietary model (GPT-4o) and are not validated against clinician-written summaries. Their quality is evidenced only indirectly through downstream recommendation performance.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Rare Medication Recommendation with Counterfactual Data Augmentation and Large Language Models." pith.science (2026). https://pith.science/paper/VS7RZYJ5

@misc{pith2026260724829,
  author       = {Pith},
  title        = {Pith review of: Improving Rare Medication Recommendation with Counterfactual Data Augmentation and Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VS7RZYJ5}},
  note         = {Machine review of arXiv:2607.24829}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

AI-based medication recommendation systems have attracted substantial attention due to their potential to enhance patient safety and therapeutic outcomes. Despite the clinical importance of accurately recommending rarely prescribed medications (rare-meds), we observe that most existing methods show significantly lower predictive performance for rare-meds. We attribute this issue to two intrinsic limitations: (a) the inherent scarcity of data for rare-meds and (b) limited consideration of co-recommended medications. To address these limitations, we propose GenRxR, a novel framework based on large language models (LLMs). GenRxR leverages the medical knowledge and clinical reasoning capability of LLMs to generate counterfactual medical data, mitigating the data scarcity issue for rare-meds. It also integrates an LLM into the medication recommendation process to model relationships among co-recommended medications. To further enhance the clinical reasoning, we introduce an instruction tuning step that aligns the LLM's capability with the recommendation task, enabling better handling of clinical context, including rare-meds cases. In our experiments, we show that GenRxR outperforms 14 (including 5 LLM-based) baselines in most cases. Specifically, it achieves up to 30.9% higher predictive performance for rare-meds than the strongest baseline.

Figures

Figures reproduced from arXiv: 2607.24829 by Buru Chang, Jaewon Kim, Kijung Shin, Shinhwan Kang, Soo Yong Lee.

Figure 1
Figure 1. Figure 1: (a) Medication Occurrence Frequency (Top) and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Example of Medical Record Summary. Outputs. The LLMCF generates the following outputs: • OCF: Generated text consisting of (1) summaries of candidate patients’ clinical contexts (medical record summary, see Fig￾ure 3) and (2) explicit rationale for selecting the most suitable patient for prescribing the target medication 𝑚∗ . • Rˆ ∗ : Generated counterfactual data for the selected patient, in￾cluding the m… view at source ↗
Figure 5
Figure 5. Figure 5: Overview of LLM-based Medication Recommendation Model [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

59 extracted references · 8 linked inside Pith

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report.arXiv:2303.08774(2023)

    Pith/arXiv arXiv 2023

  2. [2]

    Harry Amad, Zhaozhi Qian, Dennis Frauen, Julianna Piskorz, Stefan Feuerriegel, and Mihaela van der Schaar. 2025. Improving the generation and evaluation of Improving Rare Medication Recommendation RecSys ’26, September 27-October 02, 2026, Minneapolis, MN, USA synthetic data for downstream medical causal inference. InNeurIPS

    2025

  3. [3]

    Maede Ashofteh Barabadi, Xiaodan Zhu, Wai Yip Chan, Amber L Simpson, and Richard KG Do. 2025. Targeted generative data augmentation for automatic metastases detection from free-text radiology reports.Frontiers in Artificial Intelligence8 (2025), 1513674

    2025

  4. [4]

    Ananth Balashankar, Xuezhi Wang, Yao Qin, Ben Packer, Nithum Thain, Ed Chi, Jilin Chen, and Alex Beutel. 2023. Improving classifier robustness through active generative counterfactual data augmentation. InFindings of EMNLP

    2023

  5. [5]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He

  6. [6]

    Zeming Chen, Qiyue Gao, Antoine Bosselut, Ashish Sabharwal, and Kyle Richard- son. 2023. DISCO: Distilling Counterfactuals with Large Language Models. In ACL

    2023

  7. [7]

    Chanwoo Choi, Jinsoo Kim, Sukmin Cho, Soyeong Jeong, and Buru Chang. 2025. The rag paradox: A black-box attack exploiting unintentional vulnerabilities in retrieval-augmented generation systems.arXiv:2502.20995(2025)

    arXiv 2025

  8. [8]

    Edward Choi, Mohammad Taha Bahadori, Jimeng Sun, Joshua Kulas, Andy Schuetz, and Walter Stewart. 2016. Retain: An interpretable predictive model for healthcare using reverse time attention mechanism. InNeurIPS

    2016

  9. [9]

    Zeyu Cui, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. 2022. M6-rec: Generative pretrained language models are open-ended recommender systems.arXiv:2205.08084(2022)

    Pith/arXiv arXiv 2022

  10. [10]

    Sunhao Dai, Ninglu Shao, Haiyuan Zhao, Weijie Yu, Zihua Si, Chen Xu, Zhongx- iang Sun, Xiao Zhang, and Jun Xu. 2023. Uncovering chatgpt’s capabilities in recommender systems. InRecSys

    2023

  11. [11]

    Zhuyun Dai, Vincent Y Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith B Hall, and Ming-Wei Chang. 2022. Promptagator: Few-shot dense retrieval from 8 examples.arXiv:2209.11755(2022)

    Pith/arXiv arXiv 2022

  12. [12]

    Yizhou Dang, Yuting Liu, Enneng Yang, Minhan Huang, Guibing Guo, Jianzhe Zhao, and Xingwei Wang. 2025. Data augmentation as free lunch: Exploring the test-time augmentation for sequential recommendation. InSIGIR

    2025

  13. [13]

    European Medicines Agency. 2018. Orphan designation: Overview. https://www.ema.europa.eu/en/human-regulatory-overview/orphan- designation-overview. Accessed: 2025-05-16

    2018

  14. [14]

    Chenxiao Fan, Chongming Gao, Wentao Shi, Yaxin Gong, Zihao Zhao, and Fuli Feng. 2025. Fine-grained List-wise Alignment for Generative Medication Recom- mendation.arXiv:2505.20218(2025)

    Pith/arXiv arXiv 2025

  15. [15]

    Ju Fan, Tongyu Liu, Guoliang Li, Junyou Chen, Yuwei Shen, and Xiaoyong Du

  16. [16]

    Karamarie Fecho, Stanley C Ahalt, Michael Knowles, Ashok Krishnamurthy, Margaret Leigh, Kenneth Morton, Emily Pfaff, Max Wang, and Hong Yi. 2022. Leveraging open electronic health record data and environmental exposures data to derive insights into rare pulmonary disease.Frontiers in Artificial Intelligence 5 (2022), 918888

    2022

  17. [17]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5). InRecSys

    2022

  18. [18]

    Sreyan Ghosh, Utkarsh Tyagi, Sonal Kumar, and Dinesh Manocha. 2023. Bioaug: Conditional generation based data augmentation for low-resource biomedical ner. InSIGIR

    2023

  19. [19]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models.arXiv:2407.21783(2024)

    Pith/arXiv arXiv 2024

  20. [20]

    Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian McAuley, and Wayne Xin Zhao. 2024. Large language models are zero-shot rankers for recommender systems. InECIR

    2024

  21. [21]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models.. InICLR

    2022

  22. [22]

    Alistair EW Johnson, Lucas Bulgarelli, Lu Shen, Alvin Gayles, Ayad Shammout, Steven Horng, Tom J Pollard, Sicheng Hao, Benjamin Moody, Brian Gow, et al

  23. [23]

    Alistair EW Johnson, Tom J Pollard, Lu Shen, Li-wei H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G Mark. 2016. MIMIC-III, a freely accessible critical care database.Scientific data3, 1 (2016), 1–9

    2016

  24. [24]

    Shinhwan Kang, Soo Yong Lee, Jaewon Kim, Kijung Shin, and Buru Chang. 2025. Code and Datasets. https://github.com/ShinhwanKang/GenRxR

    2025

  25. [25]

    MIMIC-IV, a freely accessible electronic health record dataset.Scientific data10, 1 (2023), 1

    2023

  26. [26]

    Yongqi Li, Mayi Xu, Xin Miao, Shen Zhou, and Tieyun Qian. 2024. Prompting Large Language Models for Counterfactual Generation: An Empirical Study. In LREC-COLING

    2024

  27. [27]

    Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. 2024. Llara: Large language-recommendation assistant. In SIGIR

    2024

  28. [28]

    Tiffany H Kung, Morgan Cheatham, Arielle Medenilla, Czarina Sillos, Lorie De Leon, Camille Elepaño, Maria Madriaga, Rimel Aggabao, Giezel Diaz-Candido, James Maningo, et al. 2023. Performance of ChatGPT on USMLE: potential for AI-assisted medical education using large language models.PLoS digital health2, 2 (2023), e0000198

    2023

  29. [29]

    Qidong Liu, Xian Wu, Xiangyu Zhao, Yuanshao Zhu, Zijian Zhang, Feng Tian, and Yefeng Zheng. 2024. Large language model distilling medication recommendation model.arXiv:2402.02803(2024)

    Pith/arXiv arXiv 2024

  30. [30]

    Harsha Nori, Nicholas King, Scott Mayer McKinney, Dean Carignan, and Eric Horvitz. 2023. Capabilities of gpt-4 on medical challenge problems. arXiv:2303.13375(2023)

    Pith/arXiv arXiv 2023

  31. [31]

    Jinghui Liu, Bevan Koopman, Nathan J Brown, Kevin Chu, and Anthony Nguyen

  32. [32]

    Cheng Peng, Xi Yang, Aokun Chen, Kaleb E Smith, Nima PourNejatian, Anthony B Costa, Cheryl Martin, Mona G Flores, Ying Zhang, Tanja Magoc, et al. 2023. A study of generative large language model for medical research and healthcare. NPJ digital medicine6, 1 (2023), 210

    2023

  33. [33]

    Shauli Ravfogel, Anej Svete, Vésteinn Snæbjarnarson, and Ryan Cotterell. 2025. Gumbel Counterfactual Generation From Language Models. InICLR

    2025

  34. [34]

    AD Roberts and R. Wadhwa. 2023. Orphan Drug Approval Laws. https://www. ncbi.nlm.nih.gov/books/NBK572052/. Accessed: 2025-05-16

    2023

  35. [35]

    OpenAI. 2025. Introducing GPT-5. https://openai.com/index/introducing-gpt-5/

    2025

  36. [36]

    Junyuan Shang, Cao Xiao, Tengfei Ma, Hongyan Li, and Jimeng Sun. 2019. Gamenet: Graph augmented memory networks for recommending medication combination. InAAAI

    2019

  37. [37]

    Karan Singhal, Tao Tu, Juraj Gottweis, Rory Sayres, Ellery Wulczyn, Mohamed Amin, Le Hou, Kevin Clark, Stephen R Pfohl, Heather Cole-Lewis, et al . 2025. Toward expert-level medical question answering with large language models. Nature Medicine(2025), 1–8

    2025

  38. [38]

    Daniel Smolyak, Margrét V Bjarnadóttir, Kenyon Crowley, and Ritu Agarwal

  39. [39]

    Indira Sen, Dennis Assenmacher, Mattia Samory, Isabelle Augenstein, Wil Aalst, and Claudia Wagner. 2023. People Make Better Edits: Measuring the Efficacy of LLM-Generated Counterfactually Augmented Data for Harmful Language Detection. InEMNLP

    2023

  40. [40]

    Nicholas P Tatonetti, Patrick P Ye, Roxana Daneshjou, and Russ B Altman. 2012. Data-driven prediction of drug effects and interactions.Science translational medicine4, 125 (2012), 125ra31–125ra31

    2012

  41. [41]

    Lena Wiese and Deborah Höltje. 2021. NNCompare: a framework for dataset selection, data augmentation and comparison of different neural networks for medical image analysis. InDEEM Workshop@SIGMOD

    2021

  42. [42]

    Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, et al . 2024. A survey on large language models for recommendation.World Wide Web27, 5 (2024), 60

    2024

  43. [43]

    Rui Wu, Zhaopeng Qiu, Jiacheng Jiang, Guilin Qi, and Xian Wu. 2022. Conditional generation net for medication recommendation. InWWW

    2022

  44. [44]

    Jie Tan, Yu Rong, Kangfei Zhao, Tian Bian, Tingyang Xu, Junzhou Huang, Hong Cheng, and Helen Meng. 2024. Natural Language-Assisted Multi-modal Medica- tion Recommendation. InCIKM

    2024

  45. [45]

    Chaoqi Yang, Cao Xiao, Fenglong Ma, Lucas Glass, and Jimeng Sun. 2021. Safe- Drug: Dual Molecular Graph Encoders for Recommending Effective and Safe Drug Combinations. InIJCAI

    2021

  46. [46]

    Nianzu Yang, Kaipeng Zeng, Qitian Wu, and Junchi Yan. 2023. Molerec: Combina- torial drug recommendation with substructure-aware molecular representation learning. InWWW

    2023

  47. [47]

    Xi Yang, Aokun Chen, Nima PourNejatian, Hoo Chang Shin, Kaleb E Smith, Christopher Parisien, Colin Compas, Cheryl Martin, Anthony B Costa, Mona G Flores, et al. 2022. A large language model for electronic health records.NPJ digital medicine5, 1 (2022), 194

    2022

  48. [48]

    Han-Wook Yoo. 2023. Development of orphan drugs for rare diseases.Clinical and Experimental Pediatrics67, 7 (2023), 315

    2023

  49. [49]

    Chaoqi Yang, Cao Xiao, Lucas Glass, and Jimeng Sun. 2021. Change Matters: Medication Change Prediction with Recurrent Residual Networks. InIJCAI

    2021

  50. [50]

    Yang Zhang, Fuli Feng, Jizhi Zhang, Keqin Bao, Qifan Wang, and Xiangnan He

  51. [51]

    Zihao Zhao, Chenxiao Fan, Junlong Liu, Zheng Wang, Xiangnan He, Chong- ming Gao, Juan Li, and Fuli Feng. 2025. Fine-grained Alignment of Large Lan- guage Models for General Medication Recommendation without Overprescrip- tion.arXiv:2503.03687(2025)

    Pith/arXiv arXiv 2025

  52. [52]

    Zihao Zhao, Yi Jing, Fuli Feng, Jiancan Wu, Chongming Gao, and Xiangnan He

  53. [54]

    Yutao Zhang, Robert Chen, Jie Tang, Walter F Stewart, and Jimeng Sun. 2017. LEAP: learning to prescribe effective and safe treatment combinations for multi- morbidity. InKDD

    2017

  54. [56]

    Collm: Integrating collaborative embeddings into large language models for recommendation.IEEE TKDE37, 5 (2025), 2329–2340

    2025

  55. [59]

    Leave no patient behind: Enhancing medication recommendation for rare disease patients. InSIGIR

  56. [2020]

    Relational data synthesis using generative adversarial networks: A design space exploration.PVLDB13, 12 (2020), 1962–1975

    2020

  57. [2023]

    InRecSys

    Tallrec: An effective and efficient tuning framework to align large language model with recommendation. InRecSys

  58. [2024]

    JAMIA open7, 4 (2024), ooae114

    Large language models and synthetic health data: progress and prospects. JAMIA open7, 4 (2024), ooae114

    2024

  59. [2025]

    Generating synthetic clinical text with local large language models to identify misdiagnosed limb fractures in radiology reports.Artificial Intelligence in Medicine159 (2025), 103027

    2025

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.