KERAP: A Knowledge-Enhanced Reasoning Approach for Accurate Zero-shot Diagnosis Prediction Using Multi-agent LLMs

Carl Yang; Fadi Nahab; Hejie Cui; Jiaying Lu; Kai Shu; Xiao Hu; Yuzhang Xie; Ziyang Zhang

REVIEW 2 major objections 4 minor 1 cited by

KERAP: A Knowledge-Enhanced Reasoning Approach for Accurate Zero-shot Diagnosis Prediction Using Multi-agent LLMs

T0 review · 2 major / 4 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 The paper proposes KERAP, a multi-agent LLM framework that uses a biomedical knowledge graph to make zero-shot next-visit diagnosis predictions, and reports accuracy close to supervised models that need hundreds of thousands of labeled…

desk verdict KERAP is a well-ablated LLM+KG pipeline with a plausible zero-shot story, but the task definition never rules out target-disease leakage and the KG split produces a clinically wrong exclusion; the headline numbers need that check before they can be trusted. read the letter →

arxiv 2507.02773 v2 pith:5CHDW3SC submitted 2025-07-03 cs.AI cs.LGcs.MA

Yuzhang Xie , Hejie Cui , Ziyang Zhang , Jiaying Lu , Kai Shu , Fadi Nahab , Xiao Hu , Carl Yang This is my paper · ORCID

classification cs.AIcs.LGcs.MA

keywords zero-shotdiagnosispredictionknowledgegraphmulti-agentLLMelectronichealthrecordsentitylinkingpositiveandnegativerelationsmedicalreasoningretrieval-augmented

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

KERAP is an attempt to get reliable diagnosis prediction without any labeled training data. It splits the task among three LLM-driven agents: one links each patient attribute to an entity in a large biomedical knowledge graph, one retrieves the graph relations that count for and against the target disease, and one reasons in two stages, first for inclusion then for exclusion. On four EHR datasets, the authors report that this pipeline reaches 70-77% accuracy and F1, far above LLM prompting baselines and close to logistic regression and random forest trained on 500-1000 labeled examples. The reason to care is that many diseases lack large expert-labeled datasets, so a zero-shot predictor that leans on structured medical knowledge would make risk screening feasible where supervised training is not.

What carries the argument

The load-bearing ingredient is the knowledge graph itself, iBKH, with 2.38 million entities and 48 million relation triples, plus the way its triples are categorized. The linkage agent uses pretrained biomedical embeddings and an LLM to map a disease name from an EHR to a KG entity; the retrieval agent then splits that entity's relations into positive knowledge (inclusion criteria such as 'curcumin relates to cognitive dysfunction') and negative knowledge (exclusion criteria such as 'pravastatin does not treat cognitive dysfunction'); the prediction agent consumes these in two conversational stages and issues YES or NO. The two-stage design is what turns the KG from a text inoculation into a reasoning scaffold: Stage I forms a hypothesis from inclusion criteria, Stage II rechecks it against exclusion criteria.

What would settle it

A decisive check is to rerun KERAP on the same datasets after removing every visit whose current record already contains the target disease; if accuracy and F1 fall toward the prompting baselines, the advantage largely comes from code presence rather than knowledge-guided temporal reasoning. A second check is to have clinicians annotate a random sample of the retrieved positive and negative triples: if a substantial share of exclusion relations (for example, hypertension marked unrelated to CKD) are wrong, the guidance itself carries errors.

Watch

Extended reading notes

Core claim

The central claim is that the combination of knowledge-graph retrieval and staged multi-agent reasoning removes the two failure modes of direct LLM diagnosis prediction: hallucinated or unsupported reasoning, and a bias toward labeling almost everyone high-risk. The discovery is not a new medical fact but a pipeline finding: separating the linking, retrieval, and prediction tasks into specialized agents, and separating KG relations into positive and negative evidence, lets a general-purpose LLM match supervised classifiers in zero-shot conditions. The paper demonstrates this on four conditions (post-stroke cognitive impairment, pneumonia, chronic kidney disease, and congestive heart failure), where KERAP reports the highest accuracy and F1 among the compared zero-shot methods on all four datasets.

Load-bearing premise

The load-bearing premise is that a positive prediction is not trivially available from the target disease already being in the current visit, and that the knowledge graph's automatic relation split into inclusion and exclusion evidence is medically correct.

Editorial extensions

If this is right

Diagnosis prediction with LLMs no longer needs a labeled training set for every new disease; the same three-agent pipeline can be pointed at a new target disease as long as the knowledge graph contains it.
Adding a negative-evidence stage to a reasoning pipeline can correct the all-positive bias, since the Stage II check is aimed precisely at cases the model initially over-predicts.
Performance comparisons should start treating zero-shot pipelines as comparable to supervised classifiers, since the paper reports a gap of roughly 2% in average accuracy against logistic regression and random forest trained on 500-1000 samples.
The design separates entity linking from evidence retrieval from prediction, so each agent can be upgraded or audited independently without re-engineering the whole system.

Reading between the lines

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

The paper never states that visits where the target disease already appears in the current record were excluded, so a conservative re-test that drops such visits would show how much of the reported accuracy reflects genuine temporal generalization rather than code-presence.
Because the KG relations are classified automatically, clinically wrong labels can enter the guidance; Case B lists hypertension as unrelated to CKD, which standard medicine contradicts. Curating or validating relation polarity before retrieval could change predictions and is a cheap external check.
The same positive/negative two-stage structure could transfer to other high-stakes prediction tasks with structured prior knowledge, such as adverse drug events or hospital readmission, but the paper does not test that.

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

KERAP is a reasonable engineering contribution: it wraps SAPBERT candidate retrieval, an LLM linker, iBKH knowledge retrieval with a positive/negative split, and a two-stage self-reflection prompt into one pipeline. The ablations are the best part. They separate the KG contribution from the multi-stage contribution and show both matter, and the cost table is honest about KERAP being the most expensive method. Comparing against LR/RF at 500-1000 samples is a useful anchor.

The soft spots are real. First, the task definition never states that visits where the target disease already appears in the current record are excluded. That is the standard next-visit leak. For chronic conditions like CKD and CHF, a diagnosis code often persists, so code presence alone can predict the next positive visit. The paper's own Case A shows the predictor using 'nontraumatic subarachnoid hemorrhage' as a risk factor for PSCI, which is exactly the kind of signal that leakage would exploit. The large gap over Direct Prompting might be the KG-guided prompt amplifying that signal rather than genuine zero-shot generalization. This has to be fixed by an explicit preprocessing description and, ideally, a sensitivity analysis excluding prevalent cases.

Second, the positive/negative knowledge split is not medically reliable. Case B lists hypertension as an 'unrelated factor' for CKD, which contradicts standard clinical knowledge. That suggests the automatic relation categorization can inject false exclusions, and the retrieval agent's summaries are only as good as the KG's relation semantics. That is a correctness concern for the interpretability story.

Third, the HuatuoGPT baseline uses a different base model, so that comparison is confounded; and Figures 2-4 lack error bars even though Table 1 has them. The 'efficiently' in the abstract is also not supported: KERAP has the highest token cost and runs about equal to Iterative Prompting in time. These are minor by comparison but should be tightened.

If the leakage issue is cleared, the result is solid: a zero-shot predictor within a couple of points of supervised LR/RF is a useful finding. As written, the central claim is not yet internally validated. I'd want the authors to spell out the exclusion rule and rerun or at least analyze the prevalent cases.

This is a paper for people building LLM-based clinical prediction and for anyone designing KG-augmented reasoning pipelines. It deserves a serious referee; the concerns are addressable. Recommendation: send to peer review, but with a clear request for the prevalence-exclusion analysis.

Referee Report

2 major / 4 minor

Summary. The paper proposes KERAP, a zero-shot diagnosis prediction pipeline in which a linkage agent maps EHR concepts to entities in the iBKH knowledge graph, a retrieval agent extracts positive and negative relation knowledge, and a prediction agent performs two-stage reasoning with GPT-4o-mini. The authors evaluate on four EHR datasets (PSCI, PNA, CKD, CHF) against direct, step-by-step, KG-augmented, iterative, and pre-trained medical-LLM baselines. KERAP reports higher accuracy and F1 than all baselines on all four datasets, with ablations separating the KG and multi-stage contributions, an LLM-choice study, a cost analysis, supervised-learning comparisons, and case studies. The public repository is referenced for code, prompts, and retrieved knowledge.

Significance. If the reported results hold, the contribution is substantial: a zero-shot pipeline whose performance approaches supervised logistic regression and random forest trained on 500-1000 labeled samples, with interpretable KG-grounded reasoning, could be practically useful in low-label medical settings. The paper deserves credit for the clean architecture, the five-run averages with standard deviations in Table 1, and the explicit separation of KG and multi-stage contributions through the Iterative Prompting and KG-Augmented Prompting baselines. The central caveat is that the label-construction protocol must rule out target-disease code leakage, and the KG's positive/negative relation splits must be clinically trustworthy before the headline claim can be accepted.

major comments (2)

[Method, Task Definition] The task definition states only that the model predicts whether disease d will be diagnosed at the next visit given the current record r_i, and it does not state whether visits in which d already appears in r_i are excluded. For chronic conditions such as CKD, CHF, and PSCI, a diagnosis code once recorded often persists in subsequent visits, so code presence can almost determine the next-visit label. Case A's prompt even instructs the model to treat existing cerebrovascular diagnoses as risk factors, which would mechanically amplify such leakage. The public repository may clarify the preprocessing, but as written the paper does not rule out this failure mode. The authors should state the exclusion rule, re-run the experiments after removing target-disease codes from current-visit inputs or excluding prevalent cases, and report performance on the clean subset.
[Case Studies, Case B] The Stage II negative knowledge for CKD in Case B lists 'hypertension' as an unrelated factor, despite hypertension being a leading and well-established risk factor for CKD. This indicates that the automatic categorization of iBKH relations into RTx,+ and RTx,− can produce clinically invalid exclusion knowledge, undermining the claim that retrieved knowledge is 'clinically validated.' The example still ends with a correct prediction, so this is not by itself a demonstration of failure, but it is evidence that the relation categorization can inject errors that the framework presents as authoritative. The authors should audit the positive/negative splits for all four target diseases, quantify the frequency of clinically incorrect relations, and test the prediction agent's sensitivity to such errors.

minor comments (4)

[Figures 2 and 3] The bar charts show no error bars even though Table 1 reports standard deviations over five runs; add error bars or state explicitly that they are omitted for visual clarity.
[Comparison with Supervised Learning Methods] The text reports KERAP as 73.20% ACC and 72.88% F1, while Table 2's averages are 73.41% and 72.24%; reconcile the numbers and clarify whether Figure 4 uses the same evaluation set as Table 1.
[KG Data Source] The claim that iBKH is 'clinically validated' is stated without a citation or validation metric; please substantiate this claim or soften the wording.
[Experiments] PSCI is used both for the post-stroke cognitive impairment disease and for the dataset name; define the dataset name once to avoid ambiguity for readers.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: KERAP's predictions are generated from an external KG, a frozen LLM, and patient data with no fitting to test labels; self-citations are not load-bearing.

full rationale

KERAP's derivation chain does not reduce to its inputs. In the Method section, the linkage agent maps a disease name to an iBKH entity via SAPBERT embeddings and an LLM (Eq. 1); the retrieval agent summarizes external KG triples into positive and negative knowledge (Eq. 2); the prediction agent is a prompted, frozen GPT-4o-mini that outputs YES or NO from the patient record plus retrieved knowledge (Eq. 3). No stage fits any parameter to the target labels, and the KG and LLM are external to the paper's experiments. The reported zero-shot numbers are therefore genuinely out-of-sample with respect to training, not renamed fitted outputs. The self-citations are minor: ref. 16 (PromptLink) is cited only as inspiration for the linkage agent, not as a theorem forcing the result, and refs. 29-30 provide dataset settings. One experimental-validity concern, flagged as missing support rather than circularity, is that the Task Definition does not explicitly exclude visits where the target disease code already appears in the current record ri; if such visits were kept, code presence could leak into the next-visit label, especially for chronic diseases like CKD and CHF. This would threaten the empirical zero-shot claim but is not a circular derivation because the input and label are not defined as identical, and no fitted parameter is renamed a prediction.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

No numeric weights are fit to the evaluation data. The main free choice is LC=10 for linkage candidates. The central assumptions concern data labeling (no target leakage), KG relation validity, and LLM prompt faithfulness.

free parameters (1)

LC (linkage candidates) = 10
Number of candidate KG entities retrieved by SAPBERT cosine similarity before LLM disambiguation; chosen without reported sensitivity analysis.

assumptions (4)

domain assumption iBKH knowledge graph relations are clinically valid and useful as inclusion/exclusion criteria
Retrieval agent summarizes KG triples into positive and negative knowledge; Case B's negative knowledge lists hypertension as unrelated to CKD, which conflicts with standard medicine, showing the assumption can fail.
domain assumption GPT-4o-mini reliably executes linkage, retrieval summarization, and prediction without task-specific tuning
All three agents are prompts on the same frozen LLM; reliability is demonstrated only through two case studies, not systematic checks.
ad hoc to paper Next-visit diagnosis labels are constructed so the target disease is not already present in the current visit record
The task definition does not state this exclusion; if it is absent, the model can predict by code presence and the zero-shot claim is weakened.
domain assumption The selected EHR attributes are sufficient to represent the clinical state relevant to the target disease
Prediction depends only on recorded codes and medications; missing or incomplete records would reduce accuracy and bias comparisons.

how reviews work

0 comments

Cite this review

Pith. "Pith review of KERAP: A Knowledge-Enhanced Reasoning Approach for Accurate Zero-shot Diagnosis Prediction Using Multi-agent LLMs." pith.science (2026). https://pith.science/paper/5CHDW3SC

@misc{pith2026250702773,
  author       = {Pith},
  title        = {Pith review of: KERAP: A Knowledge-Enhanced Reasoning Approach for Accurate Zero-shot Diagnosis Prediction Using Multi-agent LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5CHDW3SC}},
  note         = {Machine review of arXiv:2507.02773}
}

read the original abstract

Medical diagnosis prediction plays a critical role in disease detection and personalized healthcare. While machine learning (ML) models have been widely adopted for this task, their reliance on supervised training limits their ability to generalize to unseen cases, particularly given the high cost of acquiring large, labeled datasets. Large language models (LLMs) have shown promise in leveraging language abilities and biomedical knowledge for diagnosis prediction. However, they often suffer from hallucinations, lack structured medical reasoning, and produce useless outputs. To address these challenges, we propose KERAP, a knowledge graph (KG)-enhanced reasoning approach that improves LLM-based diagnosis prediction through a multi-agent architecture. Our framework consists of a linkage agent for attribute mapping, a retrieval agent for structured knowledge extraction, and a prediction agent that iteratively refines diagnosis predictions. Experimental results demonstrate that KERAP enhances diagnostic reliability efficiently, offering a scalable and interpretable solution for zero-shot medical diagnosis prediction.

Figures

Figures reproduced from arXiv: 2507.02773 by the authors.

**Figure 1.** Framework of KERAP. Framework. The KERAP framework is illustrated in Fig.1. It adopts a multi-agent architecture, with GPT-4o-mini supporting each agent. The framework comprises three key components: the linkage agent, the retrieval agent, and the prediction agent. The linkage agent ALK locates the predicted disease in a biomedical KG, serving as a prerequisite for the subsequent retrieval agent to access relevant … view at source ↗

**Figure 2.** Zero-shot performance (%) of different LLM-based diagnosis prediction methods using GPT-3.5-turbo. The blue, orange, green, red, and purple bars represent the results of five compared methods: Direct Prompting, Stepby-Step Prompting, KG-Augmented Prompting, Iterative Prompting, and KERAP, respectively [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Zero-shot performance (%) of different LLM-based diagnosis prediction methods using GPT-4o. The blue, orange, green, red, and purple bars represent the results of five compared methods: Direct Prompting, Step-by-Step Prompting, KG-Augmented Prompting, Iterative Prompting, and KERAP, respectively. Comparison with Supervised Learning Methods. To further assess the performance of KERAP, we compare it with two represent… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Performance comparison of KERAP with supervised learning methods in terms of Accuracy (a) and F1- score (b). In the figures, the blue horizontal dashed line indicates the diagnosis prediction performance of KERAP, which operates without any training samples as zero-sho…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

HypKG: Hypergraph-based Knowledge Graph Contextualization for Precision Healthcare
cs.AI 2025-07 conditional novelty 6.0 of 10

HypKG integrates EHR patient context with a biomedical knowledge graph via LLM-based entity linking and a hypergraph transformer, reporting improved performance on phenotyping and post-stroke cognitive impairment prediction.

Reference graph

Works this paper leans on

31 extracted references · 29 canonical work pages · cited by 1 Pith paper

[1]

Evaluation of artificial intelligence techniques in disease diagnosis and prediction

Ghaffar Nia N, Kaplanoglu E, Nasab A. Evaluation of artificial intelligence techniques in disease diagnosis and prediction. Discover Artificial Intelligence. 2023;3(1):5

work page 2023
[2]

TACCO: Task-guided Co-clustering of Clinical Concepts and Patient Visits for Disease Subtyping based on EHR Data

Zhang Z, Cui H, Xu R, Xie Y , Ho JC, Yang C. TACCO: Task-guided Co-clustering of Clinical Concepts and Patient Visits for Disease Subtyping based on EHR Data. In: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining; 2024. p. 6324-34

work page 2024
[3]

Machine learning model to predict mental health crises from electronic health records

Garriga R, Mas J, Abraha S, Nolan J, Harrison O, Tadros G, et al. Machine learning model to predict mental health crises from electronic health records. Nature medicine. 2022;28(6):1240-8

work page 2022
[4]

Artificial intelligence in disease diagnosis: a systematic literature review, synthesizing framework and future research agenda

Kumar Y , Koul A, Singla R, Ijaz MF. Artificial intelligence in disease diagnosis: a systematic literature review, synthesizing framework and future research agenda. Journal of ambient intelligence and humanized computing. 2023;14(7):8459-86

work page 2023
[5]

Large language models encode clinical knowledge

Singhal K, Azizi S, Tu T, Mahdavi SS, Wei J, Chung HW, et al. Large language models encode clinical knowledge. Nature. 2023;620(7972):172-80

work page 2023
[6]

Preliminary analysis of the impact of lab results on large language model generated differential diagnoses

Bhasuran B, Jin Q, Xie Y , Yang C, Hanna K, Costa J, et al. Preliminary analysis of the impact of lab results on large language model generated differential diagnoses. npj Digital Medicine. 2025;8(1):166

work page 2025
[7]

Large language models for disease diagnosis: A scoping review

Zhou S, Xu Z, Zhang M, Xu C, Guo Y , Zhan Z, et al. Large language models for disease diagnosis: A scoping review. arXiv preprint arXiv:240900097. 2024

work page 2024
[8]

Llms-based few-shot disease predictions using ehr: A novel approach combining predictive agent reasoning and critical agent instruction

Cui H, Shen Z, Zhang J, Shao H, Qin L, Ho JC, et al. Llms-based few-shot disease predictions using ehr: A novel approach combining predictive agent reasoning and critical agent instruction. arXiv:240315464. 2024

work page 2024

Show all 31 references

[9]

Detecting hallucinations in large language models using semantic entropy

Farquhar S, Kossen J, Kuhn L, Gal Y . Detecting hallucinations in large language models using semantic entropy. Nature. 2024;630(8017):625-30

2024
[10]

Simulated misuse of large language models and clinical credit systems

Anibal JT, Huth HB, Gunkel J, Gregurick SK, Wood BJ. Simulated misuse of large language models and clinical credit systems. NPJ Digital Medicine. 2024;7(1):317

2024
[11]

Unifying large language models and knowledge graphs: A roadmap

Pan S, Luo L, Wang Y , Chen C, Wang J, Wu X. Unifying large language models and knowledge graphs: A roadmap. IEEE Transactions on Knowledge and Data Engineering. 2024;36(7):3580-99

2024
[12]

Large Language Models and Knowledge Graphs: Opportunities and Challenges

Pan J, Razniewski S, Kalo JC, Singhania S, Chen J, Dietze S, et al. Large Language Models and Knowledge Graphs: Opportunities and Challenges. Transactions on Graph Data and Knowledge. 2023

2023
[13]

Leveraging Medical Knowledge Graphs Into Large Language Models for Diagnosis Prediction: Design and Application Study

Gao Y , Li R, Croxford E, Caskey J, Patterson BW, Churpek M, et al. Leveraging Medical Knowledge Graphs Into Large Language Models for Diagnosis Prediction: Design and Application Study. JMIR AI. 2025;4:e58670

2025
[14]

medIKAL: Integrating Knowledge Graphs as Assistants of LLMs for Enhanced Clinical Diagnosis on EMRs

Jia M, Duan J, Song Y , Wang J. medIKAL: Integrating Knowledge Graphs as Assistants of LLMs for Enhanced Clinical Diagnosis on EMRs. In: 31st International Conference on Computational Linguistics; 2025. p. 9278-98

2025
[15]

Guiding clinical reasoning with large language models via knowledge seeds

Wu J, Wu X, Yang J. Guiding clinical reasoning with large language models via knowledge seeds. In: Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence; 2024. p. 7491-9

2024
[16]

PromptLink: Leveraging Large Language Models for Cross-Source Biomedical Concept Linking

Xie Y , Lu J, Ho J, Nahab F, Hu X, Yang C. PromptLink: Leveraging Large Language Models for Cross-Source Biomedical Concept Linking. In: Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval; 2024. p. 2589-93

2024
[17]

Chain-of-thought prompting elicits reasoning in large language models

Wei J, Wang X, Schuurmans D, Bosma M, Xia F, Chi E, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems. 2022;35:24824-37

2022
[18]

Self-reflection in llm agents: Effects on problem-solving performance

Renze M, Guven E. Self-reflection in llm agents: Effects on problem-solving performance. 2nd International Conference on Foundation and Large Language Models. 2024

2024
[19]

Evaluating the Diagnostic Performance of Large Language Models on Complex Multimodal Medical Cases

Chiu WHK, Ko WSK, Cho WCS, Hui SYJ, Chan WCL, Kuo MD. Evaluating the Diagnostic Performance of Large Language Models on Complex Multimodal Medical Cases. Journal of Medical Internet Research. 2024;26:e53724

2024
[20]

Structured clinical reasoning prompt en- hances LLM’s diagnostic capabilities in diagnosis please quiz cases

Sonoda Y , Kurokawa R, Hagiwara A, Asari Y , Fukushima T, Abe O. Structured clinical reasoning prompt en- hances LLM’s diagnostic capabilities in diagnosis please quiz cases. Japanese Journal of Radiology. 2024:1-7

2024
[21]

Huatuogpt-ii, one-stage training for medical adaption of llms

Chen J, Wang X, Ji K, Gao A, Jiang F, Chen S, et al. Huatuogpt-ii, one-stage training for medical adaption of llms. arXiv preprint arXiv:231109774. 2023

2023
[22]

RAM-EHR: Retrieval Augmentation Meets Clinical Predictions on Electronic Health Records

Xu R, Shi W, Yu Y , Zhuang Y , Jin B, Wang MD, et al. RAM-EHR: Retrieval Augmentation Meets Clinical Predictions on Electronic Health Records. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (V olume 2: Short Papers); 2024. p. 754-65

2024
[23]

RetCare: Towards Interpretable Clinical Decision Making through LLM-Driven Medical Knowledge Retrieval

Wang Z, Zhu Y , Gao J, Zheng X, Zeng Y , He Y , et al. RetCare: Towards Interpretable Clinical Decision Making through LLM-Driven Medical Knowledge Retrieval. In: Artificial Intelligence and Data Science for Healthcare: Bridging Data-Centric AI and People-Centric Healthcare; 2024

2024
[24]

Health-LLM: Personalized retrieval-augmented disease prediction system

Jin M, Yu Q, Shu D, Zhang C, Fan L, Hua W, et al. Health-LLM: Personalized retrieval-augmented disease prediction system. arXiv preprint arXiv:240200746. 2024

2024
[25]

Emerge: Integrating rag for improved multimodal ehr predictive modeling

Zhu Y , Ren C, Wang Z, Zheng X, Xie S, Feng J, et al. Emerge: Integrating rag for improved multimodal ehr predictive modeling. arXiv preprint arXiv:240600036. 2024

2024
[26]

Cpllm: Clinical prediction with large language models

Ben Shoham O, Rappoport N. Cpllm: Clinical prediction with large language models. PLOS Digital Health. 2024;3(12):e0000680

2024
[27]

Self-alignment pretraining for biomedical entity representa- tions

Liu F, Shareghi E, Meng Z, Basaldella M, Collier N. Self-alignment pretraining for biomedical entity representa- tions. arXiv preprint arXiv:201011784. 2020

2020
[28]

MIMIC-III, a freely accessible critical care database

Johnson AE, Pollard TJ, Shen L, Lehman LwH, Feng M, Ghassemi M, et al. MIMIC-III, a freely accessible critical care database. Scientific data. 2016;3(1):1-9

2016
[29]

Hypergraph transformers for ehr-based clinical predictions

Xu R, Ali MK, Ho JC, Yang C. Hypergraph transformers for ehr-based clinical predictions. AMIA Summits on Translational Science Proceedings. 2023;2023:582

2023
[30]

Predicting Post-Stroke Cognitive Impairment (PSCI) Using Multiple Machine Learning Approaches

Xie Y , Nahab F, Ge Y , Wu Y , Saurman J, Yang C, et al. Predicting Post-Stroke Cognitive Impairment (PSCI) Using Multiple Machine Learning Approaches. Stroke. 2025;56(Suppl 1):AWP175-5

2025
[31]

Biomedical discovery through the integrative biomedical knowledge hub (iBKH)

Su C, Hou Y , Zhou M, Rajendran S, Maasch JR, Abedi Z, et al. Biomedical discovery through the integrative biomedical knowledge hub (iBKH). Iscience. 2023;26(4)

2023

Pith tools

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

[1] [1]

Evaluation of artificial intelligence techniques in disease diagnosis and prediction

Ghaffar Nia N, Kaplanoglu E, Nasab A. Evaluation of artificial intelligence techniques in disease diagnosis and prediction. Discover Artificial Intelligence. 2023;3(1):5

work page 2023

[2] [2]

TACCO: Task-guided Co-clustering of Clinical Concepts and Patient Visits for Disease Subtyping based on EHR Data

Zhang Z, Cui H, Xu R, Xie Y , Ho JC, Yang C. TACCO: Task-guided Co-clustering of Clinical Concepts and Patient Visits for Disease Subtyping based on EHR Data. In: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining; 2024. p. 6324-34

work page 2024

[3] [3]

Machine learning model to predict mental health crises from electronic health records

Garriga R, Mas J, Abraha S, Nolan J, Harrison O, Tadros G, et al. Machine learning model to predict mental health crises from electronic health records. Nature medicine. 2022;28(6):1240-8

work page 2022

[4] [4]

Artificial intelligence in disease diagnosis: a systematic literature review, synthesizing framework and future research agenda

Kumar Y , Koul A, Singla R, Ijaz MF. Artificial intelligence in disease diagnosis: a systematic literature review, synthesizing framework and future research agenda. Journal of ambient intelligence and humanized computing. 2023;14(7):8459-86

work page 2023

[5] [5]

Large language models encode clinical knowledge

Singhal K, Azizi S, Tu T, Mahdavi SS, Wei J, Chung HW, et al. Large language models encode clinical knowledge. Nature. 2023;620(7972):172-80

work page 2023

[6] [6]

Preliminary analysis of the impact of lab results on large language model generated differential diagnoses

Bhasuran B, Jin Q, Xie Y , Yang C, Hanna K, Costa J, et al. Preliminary analysis of the impact of lab results on large language model generated differential diagnoses. npj Digital Medicine. 2025;8(1):166

work page 2025

[7] [7]

Large language models for disease diagnosis: A scoping review

Zhou S, Xu Z, Zhang M, Xu C, Guo Y , Zhan Z, et al. Large language models for disease diagnosis: A scoping review. arXiv preprint arXiv:240900097. 2024

work page 2024

[8] [8]

Llms-based few-shot disease predictions using ehr: A novel approach combining predictive agent reasoning and critical agent instruction

Cui H, Shen Z, Zhang J, Shao H, Qin L, Ho JC, et al. Llms-based few-shot disease predictions using ehr: A novel approach combining predictive agent reasoning and critical agent instruction. arXiv:240315464. 2024

work page 2024