ARMR: Adaptively Responsive Network for Medication Recommendation

Feiyue Wu; Shenqi Jing; Tianxing Wu

REVIEW 3 major objections 6 minor 24 references

ARMR: Adaptively Responsive Network for Medication Recommendation

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read ARMR, a medication recommendation network that treats newly introduced drugs and reused drugs differently and splits each patient's history into recent and distant segments, outperforms state-of-the-art baselines on MIMIC-III and MIMIC-IV…

desk verdict A modest empirical advance whose headline mechanism—better new-drug recommendation—is explicitly unmeasured; worth reviewing but needs a new-drug metric, Eq. (9) fixed, and the PTL split clarified. read the letter →

arxiv 2507.04428 v1 pith:WFWP2DGT submitted 2025-07-06 cs.AI cs.LG

Feiyue Wu , Tianxing Wu , Shenqi Jing This is my paper · ORCID

classification cs.AIcs.LG

keywords medicationrecommendationelectronichealthrecordspiecewisetemporallearningadaptivelyresponsivemechanismMambaMIMIC-IIIMIMIC-IVmulti-labelprediction

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that medication recommendation improves when a model explicitly separates two things previous systems blur together: drugs the patient has taken before versus drugs being introduced for the first time, and recent hospital visits versus distant ones. On MIMIC-III and MIMIC-IV, ARMR reports gains over prior state-of-the-art methods of 2.16% in Jaccard similarity and 2.55% in PRAUC, which the authors attribute to the adaptively responsive module and the piecewise temporal learning component. If true, this matters because prescribing is inherently a mixed decision—continuing what works while responding to a changed condition—and current deep learning recommenders are mostly built to maximize reuse, not to learn the pattern of new drug introduction.

What carries the argument

The load-bearing mechanism is a pair of components: (1) Piecewise Temporal Learning (PTL), which splits a visit sequence at a hyperparameter $N$ into a recent segment passed through a feedforward network and a distant segment passed through a Mamba selective state-space block, then fuses them with a cross-attention step; and (2) the Adaptively Responsive Module (ARM), which separates each historical prescription into existing drugs $m^o_i$ and new drugs $m^n_i$, encodes each stream with its own PTL, builds "used" and "unused" masks with time decay, and reweights the medication embedding table by mixing attention scores $q_o$ and $q_n$. The output is a weighted sum of a direct patient-to-drug mapping and a cosine-similarity score against the dynamic medication embeddings, trained with a blend of binary cross-entropy and multi-label hinge loss.

What would settle it

Re-run the MIMIC-III and MIMIC-IV evaluation with predictions split into drugs present in the patient's earlier visits and drugs never seen before, computing true-positive rates and precision-recall for each group separately; if a baseline such as COGNet matches or surpasses ARMR on the new-drug group while ARMR's aggregate lead comes from the old-drug group, the paper's claim that the adaptive mechanism improves new-drug introduction is falsified.

Watch

Extended reading notes

Core claim

ARMR's central claim is that the balance between reusing old medications and proposing new ones can be learned explicitly instead of left implicit, and that modeling recent and distant patient history separately produces a better temporal representation than feeding the whole sequence to one encoder. The method decomposes each past medication set into a reused subset and a newly introduced subset, encodes each with a piecewise temporal learning component, and then uses the patient's current diagnoses and procedures as a query to reweight the base medication embeddings between "used drugs" and "unused drugs" masks. The authors report consistent improvements over eight baselines on both datasets, and their ablations show that removing either the adaptive mechanism or the temporal segmentation lowers performance.

Load-bearing premise

The central claim depends on the aggregate Jaccard, F1, and PRAUC scores reflecting genuinely better medication decisions; the paper reports no drug-drug interaction safety metric and no separate measure of whether the newly introduced drugs are correct, so if the gains come from simply reusing more old drugs, the stated balance-between-continuity-and-innovation claim would not be established.

Editorial extensions

If this is right

A recommender that scores drugs differently depending on whether the patient has taken them before makes it possible to audit continuity and innovation separately, rather than only as an aggregate set accuracy.
The piecewise split with a Mamba backbone gives long patient histories a near-linear cost alternative to transformer encoders, making the temporal model feasible for patients with many visits.
If ARMR's gains replicate, methods that encode an entire medication sequence with one architecture leave performance on the table by ignoring the different roles of recent and distant records.
The case study suggests ARMR's advantage is most visible in visits where the patient's diagnoses change, because the method is designed to react to such changes by raising the score of previously unused drugs.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension the paper does not run is a per-visit breakdown of precision and recall for new versus reused drugs; such an analysis would show whether the Jaccard gain is driven by correctly adding new drugs, by extending old-drug coverage, or by both.
The same adaptive separation could be ported to other set-prediction tasks in medicine, such as lab-test ordering or procedure recommendation, where the decision also mixes continuation with innovation.
Because the "used" versus "new" split is defined purely by the patient's own history, the method's behavior on first visits is degenerate; evaluating single-visit patients separately would test whether the adaptive module helps or hurts where there is no history to exploit.
A testable hypothesis following from the paper's logic is that if a patient's diagnoses change sharply between visits, ARMR should assign noticeably higher attention to unused drugs than it does for stable patients; this can be verified by inspecting the $q_n$ weights in the trained model.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Short version: this is a modest, honest empirical paper with a real but unmeasured mechanism. If you need a strong baseline for medication recommendation on MIMIC-III/IV, ARMR is worth knowing; if you want evidence that separating new from reused drugs helps, you won't find it here.

What's new: the combination of a piecewise temporal encoder (FFN for one segment, Mamba for the other, with cross-attention) and an adaptive mask that explicitly separates newly introduced from reused medications is not in the cited literature. The ablations are designed sensibly and the gains over COGNet are consistent across two datasets with standard deviations. I believe the aggregate numbers.

Soft spots, in order of importance.

First, the paper's core motivation is untested. The abstract and introduction say the method balances reuse of historical drugs with introduction of new drugs, but Jaccard, F1, and PRAUC are computed over all 131 labels. Since roughly 70% of prescriptions are old drugs, a model could improve these metrics by getting old drugs right and ignoring new ones. The one-patient case study is not evidence. The authors themselves say in the conclusion that they plan to design quantified experiments to evaluate the ability to introduce correct new drugs. That's an explicit admission that the central claim is not yet measured.

Second, Eq. (9) has a technical flaw. hO_j is defined as a sum over visits of a time-decay term times an indicator that M_j appears anywhere in the union of historical medications. The indicator is not visit-specific, so the time-decay and frequency information in the sum collapse into a constant that does not depend on when or how often the drug was prescribed. As written, the mask only encodes whether the drug was ever used, not the recency or frequency the text claims. This needs fixing or the text clarified.

Third, the PTL segmentation appears reversed. S(1) is the earliest visit, yet the paper calls S(1..N) 'recent' and S(N+1..T) 'distant'. Unless there is an unstated reverse ordering, the FFN is applied to the distant history and Mamba to the recent one, opposite of the description. This may just be a labeling error, but it matters for interpreting the ablation results.

The absence of a DDI metric is a minor concern; not every medication recommendation paper needs one, though for clinical plausibility it would help. No code was actually verified; the link is promised but the manuscript alone can't confirm reproducibility.

Recommendation: send it out for review. The empirical work is real and the architecture is a legitimate combination, but the authors need to add a new-drug-specific metric, fix or clarify Eq. (9) and the PTL split, and ship the code. With those changes it would be a solid contribution.

Referee Report

3 major / 6 minor

Summary. The paper proposes ARMR, a medication recommendation model with two main components: a piecewise temporal learning (PTL) component that splits patient history into recent and distant segments, using a feedforward network and Mamba respectively, and an adaptively responsive module (ARM) that separates historical (reused) medications from newly introduced medications and produces dynamic medication embeddings. The model is trained with a weighted combination of binary cross-entropy and multi-label hinge loss, and evaluated on MIMIC-III and MIMIC-IV using Jaccard similarity, F1, and PRAUC. The reported results show consistent gains over the baselines, and ablations indicate that both PTL and ARM contribute to the aggregate performance. The paper claims that ARMR improves the balance between medication reuse and the introduction of new drugs.

Significance. If the central mechanism were directly validated, ARMR would be a useful contribution: it is among the first medication recommendation models to explicitly distinguish new from reused drugs, the PTL design is a reasonable way to combine short- and long-range temporal modeling, and the authors provide public code. Table 3 reports standard deviations and the ablation study is broadly consistent with the proposed design. However, the central claim about improved new-drug recommendation is not directly measured, and Eq. (9) as written does not implement the per-drug temporal decay described in the text. The significance is therefore conditional on additional stratified evaluation and on correcting the temporal-mask formulation.

major comments (3)

[§5.2, Table 3, and §6 Conclusion] The paper's advertised contribution is that ARMR improves the balance between reusing historical medications and introducing correct new drugs, but no experiment quantifies new-drug recommendation accuracy. Jaccard, F1, and PRAUC are computed over all 131 medication labels and cannot distinguish new from reused drugs; since Figure 1 indicates that roughly 70% of prescribed drugs are reused, the reported aggregate gains of 2.16% and 2.55% could be driven entirely by improved old-drug recommendations. The one-patient case study in §5.4 is not statistical evidence, and the Conclusion explicitly states that 'in the future, we plan to 1) design quantified experiments to evaluate the specific ability of all methods on introducing correct new drugs,' which is an admission that the central mechanism is currently untested. The authors should add stratified metrics separating new-drug and reused-drug performance, reported for ARMR and all baselines.
[§4.3, Eq. (9)] Equation (9) defines hO_j = Σ_{i=1}^{T-1} (α_i · e^{-β_i·(T-i)}) * 1{M_j ∈ m_h}, where m_h is the union of all historical medication sets. Because the indicator is not indexed by i, the sum factors out and hO_j is the same constant for every drug that has ever been used by the patient, regardless of when it was last prescribed or how often. This contradicts the text's claim that hO_j 'leverages the occurrence time, frequency, and time decay of medications' and undermines the per-drug temporal sensitivity of the ARM module. The correct form should be Σ_i (α_i · e^{-β_i·(T-i)}) * 1{M_j ∈ m_i}, so that recency and frequency vary per drug.
[§5.1 and Table 3] No drug-drug interaction (DDI) safety metric is reported. For a medication recommendation system that explicitly aims to introduce new drugs, DDI rate is a standard safety check used in prior work such as GAMENet, SafeDrug, and Carmen. Without it, the reported improvements leave open the possibility that the gains come at the cost of unsafe combinations. Adding a DDI metric would strengthen the paper's clinical plausibility.

minor comments (6)

[§4.1] The heading 'Overall Arcitecture' should read 'Overall Architecture.'
[§5.1] The text says 'The simpler methods (LR, ECC, LEAP) show limited effectiveness,' but ECC is neither defined nor listed in Table 3; either add ECC to the table or remove the mention.
[§5.3] The description of ARMR w/o PTL(L) is confusing: it is defined as maintaining PTL but removing the distant-history branch, yet the following sentence says the difference between ARMR w/o PTL and ARMR w/o PTL(L) is the base sequence learning architecture (RNN vs. Mamba). Please clarify the exact architectures being compared.
[§3.1] Definition 1 should read 'Electronic Health Records,' not 'Electrical Health Records.'
[Throughout] Please fix typographical errors: 'probalilities' (Section 4.6), 'emprically' (Sections 4.6 and 5.4), 'longitudial' (Section 5.2), 'avaiable' (Abstract), and 'hisotry' (Section 5.3).
[Figure 2] The x-axis is labeled 'Time Lag' without units, and the y-axis label appears to be 'Medication' rather than the plotted quantity (Jaccard similarity). Please clarify both axes.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ARMR is a standard supervised model with held-out evaluation; the unsupported new-drug claim is a measurement gap, not a circular reduction.

full rationale

ARMR is a standard supervised multi-label model trained with binary cross-entropy and multi-label hinge losses on MIMIC-III/IV and evaluated on held-out test visits. The claimed gains reported in Table 3 are direct observed differences between trained models on fixed test splits; there is no derivation in which the output is reconstructed from the model's own inputs or from fitted constants. The ARM module's used/unused masks hO/hN and the old/new sequences mo_i/mn_i are computed from the patient's previous visits (m_h = union_{k<t} m_k), not from the current visit's ground-truth medication set, so the current prediction is not definitionally contained in the mask. The loss weight alpha=0.7, inference threshold delta=0.5, and any time-decay coefficients are training/inference hyperparameters, not fitted values renamed as predictions. The paper contains no load-bearing self-citation: the cited COGNet [Wu et al., 2022] is by a different author group, and no uniqueness theorem is imported from the present authors' prior work. The skeptical concern—that the aggregate Jaccard/F1/PRAUC gains do not quantify the claimed improvement in correctly introducing new drugs, and that the Conclusion explicitly lists designing such quantified experiments as future work—is a sound evidential limitation, but it is a gap in support, not circular reasoning. Therefore the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The model relies on standard learned embeddings and hyperparameters rather than invented physical entities. The most consequential assumptions are the representativeness of the MIMIC datasets, the sufficiency of aggregate metrics without DDI safety or new-drug-specific evaluation, and the temporal semantics of the PTL split.

free parameters (4)

N (PTL split point) = not reported
The boundary between recent and distant history is a hyperparameter; the paper gives no value, no tuning procedure, and no ablation over N.
alpha_i, beta_i (time-decay controls in Eq. 9) = not reported
These control the decay of historical medication influence; the paper does not state whether they are learned, hand-set, or tuned, and provides no search range.
alpha (loss weight) = 0.7
Set empirically with no sensitivity analysis; the balance between BCE and multi-label hinge loss directly affects training.
delta (inference threshold) = 0.5
Set empirically; the binarization threshold drives the precision/recall tradeoff and is not swept or justified.

assumptions (4)

domain assumption MIMIC-III and MIMIC-IV, preprocessed to 131 ATC third-level codes following Chen et al. 2023, are a valid testbed for medication recommendation.
The whole empirical claim depends on these datasets and preprocessing choices being representative; the paper does not analyze potential data biases (e.g., critical care patients only).
domain assumption Jaccard, F1, and PRAUC are sufficient evaluation metrics for medication recommendation quality, with no safety (DDI) or novelty-specific measure.
The paper's conclusions about better medication plans are drawn exclusively from these aggregate scores.
domain assumption The input visit sequence is ordered such that the first N elements are the recent segment, or the model reverses time internally; the paper never states which.
PTL splits S into Snear = [S(1),...,S(N)] and Sfar = [S(N+1),...,S(T)] and labels them recent/distant; if the sequence is chronological this is inverted.
standard math Standard deep learning training with backpropagation, layer normalization, and attention softmax works as described.
These are routine building blocks from cited references (Ba et al. 2016, Gu and Dao 2023).

how reviews work

0 comments

Cite this review

Pith. "Pith review of ARMR: Adaptively Responsive Network for Medication Recommendation." pith.science (2026). https://pith.science/paper/WFWP2DGT

@misc{pith2026250704428,
  author       = {Pith},
  title        = {Pith review of: ARMR: Adaptively Responsive Network for Medication Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WFWP2DGT}},
  note         = {Machine review of arXiv:2507.04428}
}

read the original abstract

Medication recommendation is a crucial task in healthcare, especially for patients with complex medical conditions. However, existing methods often struggle to effectively balance the reuse of historical medications with the introduction of new drugs in response to the changing patient conditions. In order to address this challenge, we propose an Adaptively Responsive network for Medication Recommendation (ARMR), a new method which incorporates 1) a piecewise temporal learning component that distinguishes between recent and distant patient history, enabling more nuanced temporal understanding, and 2) an adaptively responsive mechanism that dynamically adjusts attention to new and existing drugs based on the patient's current health state and medication history. Experiments on the MIMIC-III and MIMIC-IV datasets indicate that ARMR has better performance compared with the state-of-the-art baselines in different evaluation metrics, which contributes to more personalized and accurate medication recommendations. The source code is publicly avaiable at: https://github.com/seucoin/armr2.

Figures

Figures reproduced from arXiv: 2507.04428 by the authors.

**Figure 2.** The similarity between two prescriptions for patients [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** The overview of ARMR’s architecture. Given the input longitudinal patient EHR which comprises M, D and P, we predict the last medication set oˆ (t) . Step 1: learning personalized medication representation E ′m, Step 2: learning the patient’s comprehensive representation hpatient, Step 3: obtaining the final predicted medication set oˆ (t) . The Mamba module is utilized in PTL, while the PTL component is used as a s… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: An example of case study. The figure shows four methods’ [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 21 canonical work pages

[1]

Deep Learn- ing for Medication Recommendation: A Systematic Sur- vey

[Ali et al., 2023] Zafar Ali, Yi Huang, Irfan Ullah, Jun- lan Feng, Chao Deng, Nimbeshaho Thierry, Asad Khan, Asim Ullah Jan, Xiaoli Shen, Wu Rui, et al. Deep Learn- ing for Medication Recommendation: A Systematic Sur- vey. Data Intelligence, 5(2):303–354,

work page 2023
[7]

A physician advi- sory system for chronic heart failure management based on knowledge patterns

[Chen et al., 2016] Zhuo Chen, Kyle Marple, Elmer Salazar, Gopal Gupta, and Lakshman Tamil. A physician advi- sory system for chronic heart failure management based on knowledge patterns. Theory and Practice of Logic Pro- gramming, 16(5-6):604–618,

work page 2016
[8]

Context-aware safe medication recom- mendations with molecular graph and ddi graph embed- ding

[Chen et al., 2023] Qianyu Chen, Xin Li, Kunnan Geng, and Mingzhong Wang. Context-aware safe medication recom- mendations with molecular graph and ddi graph embed- ding. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 7053–7060,

work page 2023
[9]

You only train once: Loss-conditional training of deep networks

[Dosovitskiy and Djolonga, 2019] Alexey Dosovitskiy and Josip Djolonga. You only train once: Loss-conditional training of deep networks. In International Conference on Learning Representations (ICLR),

work page 2019
[11]

Mamba: Linear- time sequence modeling with selective state spaces

[Gu and Dao, 2023] Albert Gu and Tri Dao. Mamba: Linear- time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752,

arXiv 2023
[12]

Attention and memory-augmented net- works for dual-view sequential learning

[He et al., 2020] Yong He, Cheng Wang, Nan Li, and Zhenyu Zeng. Attention and memory-augmented net- works for dual-view sequential learning. In Proceed- ings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD) , pages 125–134,

work page 2020
[14]

The mimic code repos- itory: enabling reproducibility in critical care research

[Johnson et al., 2018] Alistair EW Johnson, David J Stone, Leo A Celi, and Tom J Pollard. The mimic code repos- itory: enabling reproducibility in critical care research. Journal of the American Medical Informatics Association, 25(1):32–39,

work page 2018
[17]

Large language model distilling medication rec- ommendation model

[Liu et al., 2024] Qidong Liu, Xian Wu, Xiangyu Zhao, Yuanshao Zhu, Zijian Zhang, Feng Tian, and Yefeng Zheng. Large language model distilling medication rec- ommendation model. arXiv preprint arXiv:2402.02803 ,

arXiv 2024

Show all 24 references

[18]

Yet an- other rule-based approach for extracting medication infor- mation from discharge summaries

[Solt and Tikk, 2009] Ill´es Solt and Domonkos Tikk. Yet an- other rule-based approach for extracting medication infor- mation from discharge summaries. In Proceedings of the Third I2B2 Workshop on Challenges in Natural Language Processing for Clinical Data,

2009
[20]

Conditional generation net for medication recommendation

[Wu et al., 2022] Rui Wu, Zhaopeng Qiu, Jiacheng Jiang, Guilin Qi, and Xian Wu. Conditional generation net for medication recommendation. In Proceedings of the ACM Web Conference (WWW), pages 935–945,

2022
[21]

Opportunities and challenges in developing deep learning models using electronic health records data: a systematic review

[Xiao et al., 2018] Cao Xiao, Edward Choi, and Jimeng Sun. Opportunities and challenges in developing deep learning models using electronic health records data: a systematic review. Journal of the American Medical Informatics As- sociation, 25(10):1419–1428,

2018
[22]

Molerec: Combinatorial drug recom- mendation with substructure-aware molecular representa- tion learning

[Yang et al., 2023] Nianzu Yang, Kaipeng Zeng, Qitian Wu, and Junchi Yan. Molerec: Combinatorial drug recom- mendation with substructure-aware molecular representa- tion learning. In Proceedings of the ACM Web Conference (WWW), pages 4075–4085,

2023
[23]

Leap: learning to pre- scribe effective and safe treatment combinations for multi- morbidity

[Zhang et al., 2017] Yutao Zhang, Ruoqi Chen, Jie Tang, Walter F Stewart, and Jimeng Sun. Leap: learning to pre- scribe effective and safe treatment combinations for multi- morbidity. In Proceedings of the 23rd ACM SIGKDD Con- ference on Knowledge Discovery and Data Mining (KD...

2017
[24]

Knowledge- enhanced attributed multi-task learning for medicine rec- ommendation

[Zhang et al., 2023] Yingying Zhang, Xian Wu, Quan Fang, Shengsheng Qian, and Changsheng Xu. Knowledge- enhanced attributed multi-task learning for medicine rec- ommendation. ACM Transactions on Information Systems (TOIS), 41(1):1–24, 2023

2023
[2009]

4sdrug: Symptom-based set-to- set small and safe drug recommendation

[Tan et al., 2022] Yanchao Tan, Chengjun Kong, Leisheng Yu, Pan Li, Chaochao Chen, Xiaolin Zheng, Vicki S Hertzberg, and Carl Yang. 4sdrug: Symptom-based set-to- set small and safe drug recommendation. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Dis- covery a...

2022
[2012]

Layer normalization

[Ba et al., 2016] Jimmy Lei Ba, Jamie Ryan Kiros, and Ge- offrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450,

2016 arXiv
[2016]

Predicting medications from diagnostic codes with recurrent neural networks

[Bajor and Lasko, 2022] Jacek M Bajor and Thomas A Lasko. Predicting medications from diagnostic codes with recurrent neural networks. In Proceedings of the Inter- national Conference on Learning Representations (ICLR),

2022
[2017]

Dual memory neural computer for asyn- chronous two-view sequential learning

[Le et al., 2018] Hung Le, Truyen Tran, and Svetha Venkatesh. Dual memory neural computer for asyn- chronous two-view sequential learning. In Proceedings of the 24th ACM SIGKDD Conference on Knowledge Dis- covery and Data Mining (KDD), pages 1637–1645,

2018
[2018]

Learning cost-effective and interpretable treatment regimes

[Lakkaraju and Rudin, 2017] Himabindu Lakkaraju and Cynthia Rudin. Learning cost-effective and interpretable treatment regimes. In Proceedings of the Artificial Intel- ligence and Statistics (AISTATS), pages 166–175. PMLR,

2017
[2019]

Smr: medical knowledge graph embedding for safe medicine recommendation

[Gong et al., 2021] Fan Gong, Meng Wang, Haofen Wang, Sen Wang, and Mengyue Liu. Smr: medical knowledge graph embedding for safe medicine recommendation. Big Data Research, 23:100174,

2021
[2020]

Mimic-iii, a freely ac- cessible critical care database

[Johnson et al., 2016] Alistair EW Johnson, Tom J Pollard, Lu Shen, Li-wei H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo An- thony Celi, and Roger G Mark. Mimic-iii, a freely ac- cessible critical care database. Scientific Data, 3(1):1–9,

2016
[2021]

Refine: A fine-grained medication recommendation system using deep learning and person- alized drug interaction modeling

[Bhoi et al., 2023] Suman Bhoi, Mong Li Lee, Wynne Hsu, and Ngiap Chuan Tan. Refine: A fine-grained medication recommendation system using deep learning and person- alized drug interaction modeling. 36:24013–24024,

2023
[2022]

Person- alizing medication recommendation with a graph-based approach

[Bhoi et al., 2021] Suman Bhoi, Mong Li Lee, Wynne Hsu, Hao Sen Andrew Fang, and Ngiap Chuan Tan. Person- alizing medication recommendation with a graph-based approach. ACM Transactions on Information Systems (TOIS), 40(3):1–23,

2021
[2023]

Designing a pilot sequential multiple assign- ment randomized trial for developing an adaptive treat- ment strategy

[Almirall et al., 2012] Daniel Almirall, Scott N Compton, Meredith Gunlicks-Stoessel, Naihua Duan, and Susan A Murphy. Designing a pilot sequential multiple assign- ment randomized trial for developing an adaptive treat- ment strategy. Statistics in Medicine , 31(17):1887–1902,

2012

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Deep Learn- ing for Medication Recommendation: A Systematic Sur- vey

[Ali et al., 2023] Zafar Ali, Yi Huang, Irfan Ullah, Jun- lan Feng, Chao Deng, Nimbeshaho Thierry, Asad Khan, Asim Ullah Jan, Xiaoli Shen, Wu Rui, et al. Deep Learn- ing for Medication Recommendation: A Systematic Sur- vey. Data Intelligence, 5(2):303–354,

work page 2023

[2] [7]

A physician advi- sory system for chronic heart failure management based on knowledge patterns

[Chen et al., 2016] Zhuo Chen, Kyle Marple, Elmer Salazar, Gopal Gupta, and Lakshman Tamil. A physician advi- sory system for chronic heart failure management based on knowledge patterns. Theory and Practice of Logic Pro- gramming, 16(5-6):604–618,

work page 2016

[3] [8]

Context-aware safe medication recom- mendations with molecular graph and ddi graph embed- ding

[Chen et al., 2023] Qianyu Chen, Xin Li, Kunnan Geng, and Mingzhong Wang. Context-aware safe medication recom- mendations with molecular graph and ddi graph embed- ding. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 7053–7060,

work page 2023

[4] [9]

You only train once: Loss-conditional training of deep networks

[Dosovitskiy and Djolonga, 2019] Alexey Dosovitskiy and Josip Djolonga. You only train once: Loss-conditional training of deep networks. In International Conference on Learning Representations (ICLR),

work page 2019

[5] [11]

Mamba: Linear- time sequence modeling with selective state spaces

[Gu and Dao, 2023] Albert Gu and Tri Dao. Mamba: Linear- time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752,

arXiv 2023

[6] [12]

Attention and memory-augmented net- works for dual-view sequential learning

[He et al., 2020] Yong He, Cheng Wang, Nan Li, and Zhenyu Zeng. Attention and memory-augmented net- works for dual-view sequential learning. In Proceed- ings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD) , pages 125–134,

work page 2020

[7] [14]

The mimic code repos- itory: enabling reproducibility in critical care research

[Johnson et al., 2018] Alistair EW Johnson, David J Stone, Leo A Celi, and Tom J Pollard. The mimic code repos- itory: enabling reproducibility in critical care research. Journal of the American Medical Informatics Association, 25(1):32–39,

work page 2018

[8] [17]

Large language model distilling medication rec- ommendation model

[Liu et al., 2024] Qidong Liu, Xian Wu, Xiangyu Zhao, Yuanshao Zhu, Zijian Zhang, Feng Tian, and Yefeng Zheng. Large language model distilling medication rec- ommendation model. arXiv preprint arXiv:2402.02803 ,

arXiv 2024