Pith. sign in

REVIEW 1 major objections 7 minor 50 references

A single graph encoder that links disease, drug, and procedure ontologies improves EHR diagnosis prediction on MIMIC-III and MIMIC-IV.

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 ↗

LINKO integrates multiple medical ontologies with dual-axis graph propagation and LLM-based initialization, improving diagnosis prediction on MIMIC-III and MIMIC-IV.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Solid empirical EHR encoder with a credible new combination, but the unstated fold-handling for the co-occurrence graph is the one thing that could sink it. the 1 major comments →

arxiv 2508.21320 v1 pith:6LVEWIJM submitted 2025-08-29 cs.AI cs.LG

Multi-Ontology Integration with Dual-Axis Propagation for Medical Concept Representation

Mohsen Nayebi Kerdabadi , Arya Hadizadeh Moghaddam , Dongjie Wang , Zijun Yao This is my paper
classification cs.AI cs.LG
keywords medical concept representationelectronic health recordsontology graphsdual-axis message passinglarge language model initializationdiagnosis predictionrare disease predictionMIMIC-III / MIMIC-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

LINKO sets out to show that medical concept representations improve when multiple ontologies—diagnoses, drugs, procedures—are learned as one unified graph instead of separately. The framework builds a multi-level graph, initializes concept embeddings with dense vectors retrieved from a large language model prompted with ontology context, then propagates information along two axes: vertically through parent-child hierarchy and horizontally across ontologies at the same level using visit co-occurrence. On next-visit diagnosis prediction with two public EHR datasets, the authors report that LINKO beats seven prior ontology-based encoders and lifts the performance of four base sequence models. If the claim holds, the practical payoff is a plug-in concept encoder that helps EHR models most where data is scarce and codes are rare.

Core claim

The central claim is that cross-ontology edges added at every hierarchy level are what unlock better medical code embeddings. Prior encoders mostly pass messages inside one ontology or only at the leaf level; LINKO constructs horizontal co-occurrence graphs at every level, aggregates leaf co-occurrence counts upward to define parent-level edges, and couples them with vertical propagation that first moves information bottom-up through sequential adjacent-level subgraphs and then top-down with attention over ancestors. The authors argue that the combination—not any single component—produces the improvements: ablations show removing horizontal message passing causes the largest drop, and removi

What carries the argument

The central object is the Meta-KG, a multi-level heterogeneous knowledge graph that strings together separate ontology hierarchies as vertical chains and connects concepts across ontologies at each level with co-occurrence-derived edges. Three mechanisms carry the argument: graph-augmented LLM initialization (each concept embedding is a dense vector retrieved from an LLM prompted with the code, its description, and its ancestors), Horizontal Message Passing (graph attention on regular co-occurrence graphs at parent levels, optionally hypergraph attention at the leaf level), and Vertical Message Passing (bottom-up HGIP propagation over adjacent-level subgraphs, then top-down GRAM-style attent

Load-bearing premise

The load-bearing premise is that the co-occurrence counts used to build horizontal edges come only from training data; the paper computes them from visits without describing a per-fold split, so if test visits contributed to the graph, every reported improvement could be inflated.

What would settle it

Rebuild the horizontal co-occurrence adjacency matrices separately inside each training fold, using only that fold's visits, then rerun the same 5-fold comparison. If LINKO's AUPRC advantage over GRAM and HAP collapses or shrinks to noise, the reported gains come from the graph encoding test-visit information rather than from dual-axis propagation.

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

If this is right

  • Plugging LINKO into AdaCare, Transformer, RETAIN, and TCN raises AUPRC, F1, and Acc@20 on both datasets in the paper's experiments.
  • LINKO outperforms GRAM, MMORE, KAME, G-BERT, HAP, ADORE, and KAMPNet in next-visit diagnosis prediction, per the reported 5-fold means.
  • The largest label-frequency gains are in the rarest 0–25% band, and LINKO keeps an advantage as training data shrinks to 25–100%.
  • Ablations attribute the biggest drop to removing horizontal message passing entirely; removing parent-level horizontal edges hurts more than removing leaf-level ones.
  • LLM-initialized embeddings that are then refined in the graph train much faster (~60 vs ~250 epochs); freezing them degrades performance below the random-init ablation.

Where Pith is reading between the lines

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

  • If the horizontal graphs are rebuilt per training fold from scratch, the reported margins may shrink; the paper does not state that its co-occurrence counts exclude test-fold visits, so a leakage-controlled rerun is the natural next check.
  • The same dual-axis fusion recipe could transfer outside medicine to any setting with multiple taxonomies and co-occurrence observations—for example, linking product, category, and brand hierarchies from transaction logs.
  • Because parent-level horizontal edges matter more than leaf-level ones, a general design rule for heterogeneous taxonomy fusion is to densify coarse levels, where sparse child concepts can share signal through their ancestors.
  • LLM-prompted embeddings that are refined through graph propagation offer a low-hallucination way to inject external knowledge into structured prediction: use the embedding space, not the generated text.
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

1 major / 7 minor

Summary. The paper proposes LINKO, an LLM-augmented multi-ontology integration framework for medical concept representation learning in EHRs. LINKO builds a Meta-KG over ICD-9 diagnosis, ICD-9 procedure, and ATC drug ontologies, initializes node embeddings via prompted dense retrieval from GPT text-embedding-3-small, and then performs dual-axis message passing: horizontal propagation over co-occurrence-based graphs/hypergraphs at each hierarchy level, and vertical propagation via a bottom-up HGIP module followed by top-down GRAM aggregation. On MIMIC-III and MIMIC-IV, the authors evaluate next-visit diagnosis prediction with a Transformer base, comparing against GRAM, MMORE, KAME, HAP, G-BERT, ADORE, and KAMPNet, and report consistent gains in AUPRC, F1, and Acc@k, with larger improvements in rare-code bands and under data scarcity. They also report plug-in gains for AdaCare, RETAIN, TCN, and Transformer, ablations, prompt-design analysis, and a case study.

Significance. If the reported results are valid, LINKO is a useful plug-in concept encoder: it combines a clean architectural idea (cross-ontology horizontal propagation at every hierarchy level) with LLM-based initialization, and the manuscript includes several strengths — public code, means with confidence intervals, ablations that degrade sensibly, rare-code stratification, plug-in evaluation across four sequence models, and a data-scarcity analysis. The main quantitative claim, however, rests on whether the EHR co-occurrence graph is constructed in a leak-free way: the paper does not state that the graph is recomputed inside each training fold. Given that the central novelty is precisely the horizontal message passing built from visit co-occurrence, this protocol gap must be resolved before the reported superiority can be trusted.

major comments (1)
  1. [Section 2.4 and Section 3] The horizontal graph is built from the visit co-occurrence matrix Q^(L) in Eq. (2), with higher-level edges aggregated from leaf-level counts, and the leaf-level hypergraph option treats visits as hyperedges. The paper never states that Q^(L), P^(l), A_h^(l), or the hyperedges are computed using only training visits within each of the 5 folds. As written, the graph may include visits from the evaluation fold, meaning test-fold co-occurrence informs the neighbors aggregated at test time. This is transductive leakage and would inflate every reported gain, especially the rare-code bands and data-scarcity curves, and it would also distort the w/o HMP ablation that supports the central dual-axis claim. The authors must specify the per-fold construction explicitly; if the graph is built on the full data, the experiments need to be redone with fold-local co-occurrence statistics.
minor comments (7)
  1. [Section 3] The phrase "5-fold experimentation" is ambiguous. Please state whether these are cross-validation folds, how the folds are constructed, how confidence intervals are computed, and how many random seeds are used.
  2. [Section 2.4, Eq. (2); Section 3] The threshold tau(l) that defines A_h^(l) is never reported, and no sensitivity analysis is given. Also, the LLM is identified only as "GPT text-embedding-3-small" with no snapshot/date, so the initialization is not reproducible. Please report the tau values for each level/dataset and the exact model version.
  3. [Section 4.3, Table 4] The text says "when ontology integration is enabled (right-side of the table)" and "when ontology integration is removed (left-side of the table)", but the table columns are labeled "w/ Multi-level Integration" on the left and "w/o Multi-level Integration" on the right. The sides appear to be inverted in the text; please correct.
  4. [Section 2.5, Eq. (5)] The indexing in Eq. (5) is hard to follow: l and s are not clearly bound, and X^(l-s) / X^(l-s+1) appear to mix the level index and the step counter. Please rewrite with explicit indices (e.g., level L-s and L-s+1).
  5. [Table 2] In the MIMIC-III block, LINKOw/HAT is listed before LINKOw/GAT, although the text reports GAT as the better variant on MIMIC-III. Reordering the rows to match the narrative would improve readability.
  6. [References] Reference [27] is cited as "GPT-4 API", but the method uses text-embedding-3-small. Please update the reference to accurately describe the embedding model.
  7. [Figure 5] The figure caption says solid blue lines connect codes for horizontal message passing, while the color coding also uses blue for drugs. Please clarify the legend so the reader can distinguish drug nodes/edges from horizontal co-occurrence edges.

Circularity Check

0 steps flagged

No significant circularity: LINKO's gains are benchmarked against external datasets; self-citations are background, and the co-occurrence graph is a feature construction, not a renamed prediction.

full rationale

This is an empirical engineering paper, not a derivation chain. The central claim—that LINKO improves predictive encoding—is evaluated on public datasets (MIMIC-III, MIMIC-IV) against multiple published baselines, so the results are externally anchored rather than constructed from the paper's own assumptions. The horizontal co-occurrence graph (Eq. 2) is a legitimate feature transform built from within-visit code co-occurrence; it is not a quantity that is later relabeled as a prediction. The LLM initialization is a fixed external embedding step, not fitted to the prediction objective. The ablation study compares variants, and the reported improvements are empirical comparisons rather than identities. The self-citations ([12], [25], [26]) appear only as general background for EHR modeling and are not load-bearing for LINKO's novelty or its evaluation. No uniqueness theorem or ansatz is imported from the authors' prior work. The most substantive concern is a protocol gap: Section 3 does not state whether the co-occurrence matrix Q^(L) is recomputed inside each training fold, so if it were computed on the full dataset, test-fold visits could leak into the edge construction. That is a potential data-leakage / reproducibility issue, not a circularity of the kind defined here, and the paper's text does not confirm that the leak occurred. Therefore, no circular step can be exhibited, and the circularity score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central contribution is an architecture rather than a new theory, so the ledger is light. The main unstated costs are the hand-tuned co-occurrence threshold and the untested assumption that constructing co-occurrence edges from the full dataset (without per-fold separation) does not leak evaluation information into the graph.

free parameters (3)
  • co-occurrence threshold tau(l)
    Introduced in Section 2.4 to binarize conditional probabilities into horizontal graph edges; its value(s) are not reported in the experimental setup, making it a hand-tuned hyperparameter that controls graph density and, thus, the message-passing topology.
  • number of hierarchy levels L = 3
    Set to 3 in Section 3 for ICD-9 diagnosis, ICD-9 procedure, and ATC drug ontologies. The choice affects the depth of vertical propagation and the granularity of horizontal levels.
  • embedding dimension d = 256
    Shared embedding dimension for all Meta-KG nodes, stated in Section 3 Implementation. It is a standard architectural hyperparameter but is nevertheless a hand-chosen value.
axioms (4)
  • domain assumption Ontology hierarchies are tree-like and parent-child edges are clinically meaningful
    Central to vertical message passing in Section 2.5; the method assumes that general-to-specific organization in ICD and ATC is a reliable inductive bias.
  • domain assumption LLM embeddings from text-embedding-3-small encode clinically relevant semantic knowledge
    Equation (1) initializes every node embedding with an LLM output; the method assumes these vectors capture useful medical semantics that can be refined by graph layers.
  • domain assumption EHR co-occurrence is a valid proxy for cross-ontology semantic relatedness
    Section 2.4 constructs all horizontal edges from visit-level co-occurrence counts; the validity of this proxy is assumed rather than tested.
  • standard math GAT, HAT, and GRAM are standard and correctly implemented
    The graph operators used in Eqs. (3)-(7) are taken as reliable black-box tools from the literature; no correctness proof is provided.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Ontology Integration with Dual-Axis Propagation for Medical Concept Representation." pith.science (2026). https://pith.science/paper/6LVEWIJM

@misc{pith2026250821320,
  author       = {Pith},
  title        = {Pith review of: Multi-Ontology Integration with Dual-Axis Propagation for Medical Concept Representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6LVEWIJM}},
  note         = {Machine review of arXiv:2508.21320}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Medical ontology graphs map external knowledge to medical codes in electronic health records via structured relationships. By leveraging domain-approved connections (e.g., parent-child), predictive models can generate richer medical concept representations by incorporating contextual information from related concepts. However, existing literature primarily focuses on incorporating domain knowledge from a single ontology system, or from multiple ontology systems (e.g., diseases, drugs, and procedures) in isolation, without integrating them into a unified learning structure. Consequently, concept representation learning often remains limited to intra-ontology relationships, overlooking cross-ontology connections. In this paper, we propose LINKO, a large language model (LLM)-augmented integrative ontology learning framework that leverages multiple ontology graphs simultaneously by enabling dual-axis knowledge propagation both within and across heterogeneous ontology systems to enhance medical concept representation learning. Specifically, LINKO first employs LLMs to provide a graph-retrieval-augmented initialization for ontology concept embedding, through an engineered prompt that includes concept descriptions, and is further augmented with ontology context. Second, our method jointly learns the medical concepts in diverse ontology graphs by performing knowledge propagation in two axes: (1) intra-ontology vertical propagation across hierarchical ontology levels and (2) inter-ontology horizontal propagation within every level in parallel. Last, through extensive experiments on two public datasets, we validate the superior performance of LINKO over state-of-the-art baselines. As a plug-in encoder compatible with existing EHR predictive models, LINKO further demonstrates enhanced robustness in scenarios involving limited data availability and rare disease prediction.

Figures

Figures reproduced from arXiv: 2508.21320 by Arya Hadizadeh Moghaddam, Dongjie Wang, Mohsen Nayebi Kerdabadi, Zijun Yao.

Figure 1
Figure 1. Figure 1: LINKO has three steps: (1) Meta-KG construction, which integrates heterogeneous medical concepts from multiple ontologies and initializes their embeddings using graph-augmented LLM dense vectors. The Meta-KG supports Horizontal (co-occurrence) and Vertical (hierarchical) Message Passing. (2) Horizontal Message Passing (HMP), which links concepts across ontologies at corresponding hierarchy levels based on … view at source ↗
Figure 2
Figure 2. Figure 2: Performance enhancement evaluation before and after integrating [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Performance comparison of integration of ontology encoder baselines into the base model (Transformer) across [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance evaluation across different training set sizes using the MIMIC-III and MIMIC-IV datasets. An asterisk (*) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Case study: Example of an extracted sub-KG for [PITH_FULL_IMAGE:figures/full_fig_p009_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

50 extracted references · 38 canonical work pages · 4 internal anchors

  1. [1]

    Badr AlKhamissi, Millicent Li, Asli Celikyilmaz, Mona Diab, and Marjan Ghazvininejad. 2022. A review on language models as knowledge bases. arXiv preprint arXiv:2204.06031 (2022)

    Pith/arXiv arXiv 2022

  2. [2]

    Yang An, Haocheng Tang, Bo Jin, Yi Xu, and Xiaopeng Wei. 2023. KAMPNet: multi-source medical knowledge augmented medication prediction network with multi-level graph contrastive learning. BMC Medical Informatics and Decision Making 23, 1 (2023), 243

    work page 2023

  3. [3]

    Awais Ashfaq, Anita Sant’Anna, Markus Lingman, and Sławomir Nowaczyk

    work page

  4. [4]

    Shaojie Bai, J Zico Kolter, and Vladlen Koltun. 2018. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv. arXiv preprint arXiv:1803.01271 10 (2018)

    Pith/arXiv arXiv 2018

  5. [5]

    Song Bai, Feihu Zhang, and Philip HS Torr. 2021. Hypergraph convolution and hypergraph attention. Pattern Recognition 110 (2021), 107637

    2021

  6. [6]

    Derun Cai, Chenxi Sun, Moxian Song, Baofeng Zhang, Shenda Hong, and Hongyan Li. 2022. Hypergraph contrastive learning for electronic health records. In Proceedings of the 2022 SIAM International Conference on Data Mining (SDM) . SIAM, 127–135

    work page 2022

  7. [7]

    Chin Wang Cheong, Kejing Yin, William K Cheung, Benjamin CM Fung, and Jonathan Poon. 2023. Adaptive Integration of Categorical and Multi-relational Ontologies with EHR Data for Medical Concept Embedding. ACM Transactions on Intelligent Systems and Technology 14, 6 (2023), 1–20

    work page 2023

  8. [8]

    Edward Choi, Mohammad Taha Bahadori, Andy Schuetz, Walter F Stewart, and Jimeng Sun. 2016. Doctor ai: Predicting clinical events via recurrent neural networks. In Machine learning for healthcare conference . PMLR, 301–318

    work page 2016

  9. [9]

    Edward Choi, Mohammad Taha Bahadori, Le Song, Walter F Stewart, and Jimeng Sun. 2017. GRAM: graph-based attention model for healthcare representation learning. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining . 787–795

    work page 2017

  10. [10]

    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. Advances in neural information processing systems 29 (2016)

    work page 2016

  11. [11]

    Edward Choi, Zhen Xu, Yujia Li, Michael Dusenberry, Gerardo Flores, Emily Xue, and Andrew Dai. 2020. Learning the graphical structure of electronic health records with graph convolutional transformer. In Proceedings of the AAAI conference on artificial intelligence , Vol. 34. 606–613

    work page 2020

  12. [12]

    Arya Hadizadeh Moghaddam, Mohsen Nayebi Kerdabadi, Bin Liu, Mei Liu, and Zijun Yao. 2025. Discovering Time-aware Hidden Dependencies with Person- alized Graphical Structure in Electronic Health Records. ACM Transactions on Knowledge Discovery from Data 19, 2 (2025), 1–21

    work page 2025

  13. [13]

    Jinxiang Hu, Mohsen Nayebi Kerdabadi, Xiaohang Mei, Joseph Cappelleri, Richard Barohn, and Zijun Yao. 2025. Recurrent neural networks and attention scores for personalized prediction and interpretation of patient-reported outcomes. Journal of Biopharmaceutical Statistics (2025), 1–11

    work page 2025

  14. [14]

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2025. A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions. ACM Trans. Inf. Syst. 43, 2, Article 42 (Jan. 2025), 55 pages. https://doi.org/10.1145/3703155

    doi:10.1145/3703155 2025

  15. [15]

    Pengcheng Jiang, Cao Xiao, Adam Cross, and Jimeng Sun. 2023. Graphcare: Enhancing healthcare predictions with personalized knowledge graphs. arXiv preprint arXiv:2305.12788 (2023)

    Pith/arXiv arXiv 2023

  16. [16]

    Pengcheng Jiang, Cao Xiao, Minhao Jiang, Parminder Bhatia, Taha Kass-Hout, Jimeng Sun, and Jiawei Han. 2024. Reasoning-Enhanced Healthcare Predictions with Knowledge Graph Community Retrieval. arXiv preprint arXiv:2410.04585 (2024)

    Pith/arXiv arXiv 2024

  17. [17]

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

    work page

  18. [18]

    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 data 3, 1 (2016), 1–9

    work page 2016

  19. [19]

    Yikuan Li, Shishir Rao, José Roberto Ayala Solares, Abdelaali Hassaine, Rema Ramakrishnan, Dexter Canoy, Yajie Zhu, Kazem Rahimi, and Gholamreza Salimi- Khorshidi. 2020. BEHRT: transformer for electronic health records. Scientific reports 10, 1 (2020), 7155

    work page 2020

  20. [20]

    Chang Lu, Chandan K Reddy, Prithwish Chakraborty, Samantha Kleinberg, and Yue Ning. 2021. Collaborative graph learning with auxiliary text for temporal event prediction in healthcare. arXiv preprint arXiv:2105.07542 (2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  21. [21]

    Chang Lu, Chandan K Reddy, and Yue Ning. 2021. Self-supervised graph learning with hyperbolic embedding for temporal health event prediction. IEEE Transac- tions on Cybernetics 53, 4 (2021), 2124–2136

    work page 2021

  22. [22]

    Yuxing Lu, Xukai Zhao, and Jinzhuo Wang. 2024. ClinicalRAG: Enhancing Clinical Decision Support through Heterogeneous Knowledge Retrieval. In Proceedings of the 1st Workshop on Towards Knowledgeable Language Models (KnowLLM 2024) . 64–68

    work page 2024

  23. [23]

    Fenglong Ma, Quanzeng You, Houping Xiao, Radha Chitta, Jing Zhou, and Jing Gao. 2018. Kame: Knowledge-based attention model for diagnosis prediction in healthcare. In Proceedings of the 27th ACM international conference on information and knowledge management. 743–752

    work page 2018

  24. [24]

    Liantao Ma, Junyi Gao, Yasha Wang, Chaohe Zhang, Jiangtao Wang, Wenjie Ruan, Wen Tang, Xin Gao, and Xinyu Ma. 2020. Adacare: Explainable clinical health status representation learning via scale-adaptive feature extraction and recalibration. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 34. 825–832

    work page 2020

  25. [25]

    Arya Hadizadeh Moghaddam, Mohsen Nayebi Kerdabadi, Mei Liu, and Zijun Yao. 2024. Contrastive Learning on Medical Intents for Sequential Prescription Recommendation. arXiv preprint arXiv:2408.10259 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  26. [26]

    Mohsen Nayebi Kerdabadi, Arya Hadizadeh Moghaddam, Bin Liu, Mei Liu, and Zijun Yao. 2023. Contrastive learning of temporal distinctiveness for survival analysis in electronic health records. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management . 1897–1906

    work page 2023

  27. [27]

    OpenAI. 2023. GPT-4 API. https://platform.openai.com

    work page 2023

  28. [28]

    Judea Pearl. 2022. Reverend Bayes on inference engines: A distributed hierarchical approach. In Probabilistic and causal inference: the works of Judea Pearl . 129–138

    work page 2022

  29. [29]

    Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. 2019. Language models as knowledge bases? arXiv preprint arXiv:1909.01066 (2019)

    Pith/arXiv arXiv 2019

  30. [30]

    Raphael Poulain and Rahmatollah Beheshti. 2024. Graph transformers on EHRs: Better representation improves downstream performance. In The Twelfth Inter- national Conference on Learning Representations

    work page 2024

  31. [31]

    Fatemeh Zahra Safaeipour and Morteza Hashemi. 2024. Semantic-aware and goal-oriented communications for object detection in wireless end-to-end image transmission. arXiv preprint arXiv:2402.01064 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  32. [32]

    Junyuan Shang, Tengfei Ma, Cao Xiao, and Jimeng Sun. 2019. Pre-training of graph augmented transformers for medication recommendation. arXiv preprint arXiv:1906.00346 (2019)

    Pith/arXiv arXiv 2019

  33. [33]

    Lihong Song, Chin Wang Cheong, Kejing Yin, William K Cheung, Benjamin CM Fung, and Jonathan Poon. 2019. Medical Concept Embedding with Multiple Ontological Representations.. In IJCAI, Vol. 19. 4613–4619

    work page 2019

  34. [34]

    Chenhao Su, Sheng Gao, and Si Li. 2020. GATE: graph-attention augmented temporal neural network for medication recommendation. IEEE Access 8 (2020), 125447–125458

    work page 2020

  35. [35]

    A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems (2017)

    2017

  36. [36]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2017. Graph attention networks. arXiv preprint arXiv:1710.10903 (2017)

    Pith/arXiv arXiv 2017

  37. [37]

    World Health Organization. 2025. International Classification of Diseases (ICD). https://www.who.int/standards/classifications/classification-of-diseases. WHO Family of International Classifications

    work page 2025

  38. [38]

    World Health Organization Collaborating Centre for Drug Statistics Methodology

    work page

  39. [39]

    Ran Xu, Mohammed K Ali, Joyce C Ho, and Carl Yang. 2023. Hypergraph transformers for ehr-based clinical predictions. AMIA Summits on Translational Science Proceedings 2023 (2023), 582

    work page 2023

  40. [40]

    Ran Xu, Wenqi Shi, Yue Yu, Yuchen Zhuang, Bowen Jin, May D Wang, Joyce C Ho, and Carl Yang. 2024. Ram-ehr: Retrieval augmentation meets clinical predictions on electronic health records. arXiv preprint arXiv:2403.00815 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  41. [41]

    Ran Xu, Yue Yu, Chao Zhang, Mohammed K Ali, Joyce C Ho, and Carl Yang. 2022. Counterfactual and factual reasoning over hypergraphs for interpretable clinical predictions on ehr. In Machine Learning for Health . PMLR, 259–278

    work page 2022

  42. [42]

    Nianzu Yang, Kaipeng Zeng, Qitian Wu, and Junchi Yan. 2023. Molerec: Combina- torial drug recommendation with substructure-aware molecular representation learning. In Proceedings of the ACM Web Conference 2023 . 4075–4085

    work page 2023

  43. [43]

    Muchao Ye, Suhan Cui, Yaqing Wang, Junyu Luo, Cao Xiao, and Fenglong Ma

    work page

  44. [44]

    Muhan Zhang, Christopher R King, Michael Avidan, and Yixin Chen. 2020. Hierar- chical attention propagation for healthcare representation learning. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Multi-Ontology Integration with Dual-Axis Propagation for Medical Concept Representation Conference’17, July 2017, Wa...

    work page 2020

  45. [45]

    Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, et al . 2023. Siren’s song in the AI ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219 (2023)

    Pith/arXiv arXiv 2023

  46. [46]

    Yinghao Zhu, Changyu Ren, Zixiang Wang, Xiaochen Zheng, Shiyun Xie, Junlan Feng, Xi Zhu, Zhoujun Li, Liantao Ma, and Chengwei Pan. 2024. EMERGE: Enhancing Multimodal Electronic Health Records Predictive Modeling with Retrieval-Augmented Generation. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management (Boise, ID, ...

    arXiv 2024

  47. [2019]

    Journal of biomedical informatics 97 (2019), 103256

    Readmission prediction using deep learning on electronic health records. Journal of biomedical informatics 97 (2019), 103256

    work page 2019

  48. [2021]

    In Proceedings of the Web Conference 2021

    Medpath: Augmenting health risk prediction via medical knowledge paths. In Proceedings of the Web Conference 2021 . 1397–1409

    work page 2021

  49. [2023]

    Scientific data 10, 1 (2023), 1

    MIMIC-IV, a freely accessible electronic health record dataset. Scientific data 10, 1 (2023), 1

    work page 2023

  50. [2025]

    https: //www.whocc.no/

    Anatomical Therapeutic Chemical (ATC) Classification System. https: //www.whocc.no/. WHOCC; first published 1976

    work page 1976

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.