Paper deep dive
MED-COPILOT: A Medical Assistant Powered by GraphRAG and Similar Patient Case Retrieval
Shuheng Chen, Namratha Patil, Haonan Pan, Angel Hsing-Chi Hwang, Yao Du, Ruishan Liu, Jieyu Zhao
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 7/20/2026, 5:44:08 AM
Summary
The paper introduces MED-COPILOT, an interactive clinical decision-support system that integrates GraphRAG retrieval from WHO and NICE guidelines with hybrid semantic-keyword retrieval of similar patient cases from MIMIC-IV and Synthea. The system aims to improve clinical reasoning accuracy and generation fidelity by combining structured guideline knowledge with patient-level analogical evidence, outperforming standard RAG and parametric LLM baselines.
Entities (14)
Relation Signals (14)
MED-COPILOT → buildsknowledgegraphfrom → WHO
confidence 95% · builds a structured knowledge graph from WHO and NICE guidelines
MED-COPILOT → buildsknowledgegraphfrom → NICE
confidence 95% · builds a structured knowledge graph from WHO and NICE guidelines
MED-COPILOT → uses → GraphRAG
confidence 95% · combines guideline-grounded GraphRAG retrieval with hybrid semantic-keyword similar-patient retrieval
MED-COPILOT → uses → Similar-Patient Retrieval
confidence 95% · combines guideline-grounded GraphRAG retrieval with hybrid semantic-keyword similar-patient retrieval
MED-COPILOT → usesdatafrom → MIMIC-IV
confidence 95% · maintains a 36,000-case similar-patient database derived from SOAP-normalized MIMIC-IV notes
MED-COPILOT → usesdatafrom → Synthea
confidence 95% · derived from SOAP-normalized MIMIC-IV notes and Synthea-generated records
MED-COPILOT → evaluatedon → MIMIC-IV
confidence 90% · evaluate our framework on clinical note completion... on MIMIC-IV
MED-COPILOT → evaluatedon → MMLU
confidence 90% · evaluate our framework on... MMLU
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Clinical decision-making requires synthesizing heterogeneous evidence, including patient histories, clinical guidelines, and trajectories of comparable cases. While large language models (LLMs) offer strong reasoning capabilities, they remain prone to hallucinations and struggle to integrate long, structured medical documents. We present MED-COPILOT, an interactive clinical decision-support system designed for clinicians and medical trainees, which combines guideline-grounded GraphRAG retrieval with hybrid semantic-keyword similar-patient retrieval to support transparent and evidence-aware clinical reasoning. The system builds a structured knowledge graph from WHO and NICE guidelines, applies community-level summarization for efficient retrieval, and maintains a 36,000-case similar-patient database derived from SOAP-normalized MIMIC-IV notes and Synthea-generated records. We evaluate our framework on clinical note completion and medical question answering, and demonstrate that it consistently outperforms parametric LLM baselines and standard RAG, improving both generation fidelity and clinical reasoning accuracy. The full system is available at this https URL , enabling users to inspect retrieved evidence, visualize token-level similarity contributions, and conduct guided follow-up analysis. Our results demonstrate a practical and interpretable approach to integrating structured guideline knowledge with patient-level analogical evidence for clinical LLMs.
Tags
Links
- Source: https://arxiv.org/abs/2603.00460v1
- Canonical: https://arxiv.org/abs/2603.00460v1
Trouble viewing inline? Open PDF directly →
Full Text
33,393 characters extracted from source content.
Expand or collapse full text
MED-COPILOT: A Medical Assistant Powered by GraphRAG and Similar Patient Case Retrieval Shuheng Chen, Namratha Patil, Haonan Pan, Angel Hsing-Chi Hwang, Yao Du, Ruishan Liu, Jieyu Zhao University of Southern California Abstract Clinical decision-making requires synthesizing heterogeneous evidence, including patient his- tories, clinical guidelines, and trajectories of comparable cases. While large language mod- els (LLMs) offer strong reasoning capabilities, they remain prone to hallucinations and strug- gle to integrate long, structured medical docu- ments. We present MED-COPILOT, an interac- tive clinical decision-support system designed for clinicians and medical trainees, which com- bines guideline-grounded GraphRAG retrieval with hybrid semantic–keyword similar-patient retrieval to support transparent and evidence- aware clinical reasoning. The system builds a structured knowledge graph from WHO and NICE guidelines, applies community-level summarization for efficient retrieval, and main- tains a 36,000-case similar-patient database de- rived from SOAP-normalized MIMIC-IV notes and Synthea-generated records. We evaluate our framework on clinical note completion and medical question answering, and demonstrate that it consistently outper- forms parametric LLM baselines and stan- dard RAG, improving both generation fidelity and clinical reasoning accuracy. The full sys- tem is available athttps://huggingface.co/ spaces/Cryo3978/Med_GraphRAG, enabling users to inspect retrieved evidence, visual- ize token-level similarity contributions, and conduct guided follow-up analysis. Our re- sults demonstrate a practical and interpretable approach to integrating structured guideline knowledge with patient-level analogical evi- dence for clinical LLMs. 1 Introduction Clinical reasoning requires integrating heteroge- neous evidence sources, including patient histories, population-level clinical guidelines, and outcomes from comparable prior cases (Esteva et al., 2019; Figure 1: Overview of the proposed system, which per- forms dual evidence retrieval from similar patient cases and graph-structured clinical guidelines to support clini- cal inference. Rajkomar et al., 2019). While large language mod- els (LLMs) exhibit strong open-domain reasoning capabilities, their use in medical settings remains risky due to hallucinations, confidence miscalibra- tion, and a lack of verifiable evidence grounding (Pandit et al., 2025; Kim et al., 2025; Huang et al., 2025). Consequently, there is growing consensus that safe and trustworthy medical AI must rely on explicit retrieval of external clinical evidence rather than purely parametric knowledge (Nazi and Peng, 2024; Asgari et al., 2025), particularly for clinical decision-support systems intended to assist clini- cians in real-world reasoning workflows. Retrieval-Augmented Generation (RAG) frame- works partially address this need by grounding model outputs in retrieved text (Lewis et al., 2020; 1 arXiv:2603.00460v1 [cs.AI] 28 Feb 2026 Izacard and Grave, 2021). However, conventional RAG operates over flat, unstructured corpora and struggles to model the relational and hierarchical structure essential for clinical reasoning (Lewis et al., 2020; Zhang et al., 2025). Recent work has therefore emphasized graph-based retrieval, such as knowledge-graph-augmented RAG and GraphRAG, to support multi-step reasoning over structured clinical knowledge (Liang et al., 2025; Sanmartin, 2024; Wu et al., 2025). Despite these advances, a critical aspect of clini- cal reasoning remains largely unsupported: patient- level analogical reasoning. In clinical practice, de- cisions are frequently informed by comparisons to clinically similar patients with shared symptoms, comorbidities, or disease trajectories (Beam and Kohane, 2018; Rajkomar et al., 2018). Existing clinical RAG systems primarily focus on guide- line or document retrieval and rarely incorporate patient-level comparators into the reasoning loop (Wu et al., 2025; Zhang et al., 2025; Lu et al., 2024), leaving an important gap in case-centric decision. To bridge these limitations, we propose a uni- fied and interpretable framework that integrates guideline-grounded GraphRAG retrieval, hybrid similar-patient retrieval, and LLM-based reasoning within a single pipeline. The system constructs a structured knowledge graph from authoritative WHO and NICE clinical guidelines (Organization, 2024; for Health and Excellence, 2024) and pairs it with a large-scale similar-patient database derived from SOAP-normalized MIMIC-IV ICU records and Synthea-generated synthetic trajectories (John- son et al., 2023; Walonoski et al., 2018). A hybrid similarity function combines clinically weighted keyword matching with semantic embedding sim- ilarity, enabling flexible and transparent evidence selection during inference. We evaluate our framework on both generative and discriminative medical reasoning tasks, includ- ing clinical note completion on MIMIC-IV and standardized medical question answering bench- marks such as MedQA and the clinical subset of MMLU (Jin et al., 2021; Hendrycks et al., 2021). Across settings, the proposed approach consistently outperforms LLM-only and conventional RAG baselines, demonstrating the complementary value of structured guideline knowledge and patient-level analogical evidence. We additionally release an in- teractive system to support evidence inspection and reproducible clinical reasoning workflows. Our contributions are threefold: (1) we introduce the first unified framework that jointly incorporates graph-structured guideline retrieval and similar- patient retrieval for clinical reasoning; (2) we con- struct scalable and standardized clinical evidence resources spanning both guidelines and patient tra- jectories; and (3) we demonstrate consistent per- formance gains across multiple medical reasoning tasks with a publicly available interactive system. 2 Related Work 2.1 Retrieval-Augmented Clinical Reasoning Retrieval-Augmented Generation (RAG) enhances large language models by grounding generation in externally retrieved evidence, and has become a widely adopted paradigm for medical question an- swering and clinical reasoning (Lewis et al., 2020; Izacard and Grave, 2021; Xiong et al., 2024a). Domain-specific variants adapt RAG to biomed- ical corpora such as PubMed, clinical guidelines, and electronic health records, demonstrating sub- stantial improvements over LLM-only or chain-of- thought baselines on different benchmarks includ- ing MedQA (Xiong et al., 2024a,b). However, prior studies consistently report that text-based RAG remains sensitive to retrieval noise and degrades under long-context and multi-hop reasoning scenarios, limiting its reliability for com- plex clinical decision-making (Tang and Yang, 2024; Sun et al., 2025; Gupta et al., 2024). To address these limitations, recent work incorpo- rates structured knowledge graphs into retrieval pipelines, enabling explicit modeling of biomedical entities and relations for more interpretable multi- step reasoning (Yasunaga et al., 2021; Sohn et al., 2025). GraphRAG-style systems further organize retrieved evidence into graph-structured representa- tions and community-level summaries, improving provenance tracking and reasoning consistency in medical QA and diagnostic tasks (Wu et al., 2025; Sekar et al., 2025; Yang et al., 2024). Despite these advances, existing systems primarily focus on guideline- or document-level evidence and largely overlook patient-level analogical reasoning. 2.2 Similar-Patient Retrieval Similar-patient retrieval has been extensively stud- ied in intensive care analytics and personalized risk modeling. Early work demonstrates that trajectory- based similarity enables cohort-level outcome esti- mation and interpretable subgroup analysis in ICU populations (Lee et al., 2015; Alcaide and Aerts, 2 2021). More recent representation-learning ap- proaches model temporal dynamics and relational structures among patients to improve predictive performance (Sun et al., 2022; Ma et al., 2023). Despite its clinical relevance, prior work typi- cally treats patient similarity as a standalone an- alytical or predictive tool, rather than integrating similar-patient evidence into retrieval-augmented or LLM-based reasoning pipelines (Sharafoddini et al., 2017; Zhao et al., 2023).As a result, patient-level analogical evidence remains underuti- lized in current clinical RAG systems, motivating the urgent need for unified frameworks that com- bine structured guideline knowledge with clinically comparable patient trajectories. 3 Methodology This section describes the design of a unified re- trieval framework that integrates graph-augmented clinical guidelines and similar-patient evidence. We focus on how authoritative medical knowledge is represented, indexed, and retrieved to support interactive and transparent clinical reasoning. 3.1 System Overview The proposed system supports clinical queries in the form of either SOAP-formatted patient cases or free-text questions. Given an input, it retrieves two complementary forms of evidence: (i) graph- structured guideline knowledge encoding multi- step clinical logic, and (i) clinically similar patient cases capturing trajectory-level analogical patterns. Retrieved evidence is aggregated and presented with explicit provenance to support downstream reasoning and user inspection. 3.2 Graph-Augmented Guideline Retrieval Guideline Sources. To ensure evidential relia- bility and clinical safety, all guideline knowledge is sourced exclusively from internationally sanc- tioned authorities. Specifically, we curate clinical practice guidelines from the World Health Organi- zation (WHO) and the National Institute for Health and Care Excellence (NICE), both of which pro- vide consensus-based, evidence-graded recommen- dations with transparent revision and disclosure procedures (Organization, 2024; for Health and Excellence, 2024). After normalization and de-duplication, the re- sulting corpus comprises 118 WHO guidelines and 525 NICE guidelines, totaling over 40 million tokens of clinically actionable recommendations. These guidelines explicitly articulate eligibility cri- teria, contraindications, risk boundaries, and esca- lation pathways, forming a reliable substrate for structured retrieval and multi-hop reasoning. Graph Construction. Clinical guidelines are long, hierarchically structured documents whose logic cannot be fully captured through flat text seg- mentation alone. While segmenting guidelines into semantically coherent units supports fine-grained retrieval, it fails to expose relational structures such as conditional applicability, treatment dependen- cies, and escalation logic. To address this limitation, each guideline is first segmented into semantically coherent TextUnits, where each unit corresponds to a single recom- mendation, decision step, or well-defined clinical scenario. Segmentation respects the native guide- line hierarchy rather than arbitrary token windows. Each TextUnit is annotated with metadata derived from the original guideline document, including its source, section title, and character span, enabling transparent provenance tracking and citation. We then construct a clinical knowledge graph by extracting clinically meaningful entities (e.g., diseases, symptoms, medications, procedures) and relations (e.g., indication, contraindication, mon- itoring, escalation) from the segmented TextU- nits. Entity mentions are identified using medi- cal named-entity recognition and domain-adapted LLM prompting, and normalized to standard biomedical vocabularies (e.g., SNOMED-CT, ICD- 10, ATC). Relations and constraint qualifiers (such as age limits or escalation triggers) are encoded ex- plicitly, and all graph elements retain links to their originating guideline sections. This graph represen- tation enables structured navigation of guideline logic beyond only what is accessible through text- only retrieval. Indexing. To support efficient retrieval across both fine-grained recommendations and higher- level guideline logic, we adopt a hybrid indexing strategy. Guideline TextUnits, graph community summaries, and entity descriptions are embedded using BioClinicalBERT and stored in a local vec- tor database (LanceDB). Indexing is organized by evidence modality, preserving the structural hierar- chy of the guideline graph while enabling flexible retrieval at different levels of granularity with full provenance traceability. 3 3.3 Similar-Patient Retrieval Patient Data Sources. Similar-patient retrieval operates over a curated repository of 36,000 struc- tured patient cases. To balance clinical realism and controlled coverage, the repository integrates two complementary sources. First, 18,000 intensive care encounters are sampled from MIMIC-IV, a large-scale de-identified critical care database con- taining comprehensive EHR-derived clinical docu- mentation (Johnson et al., 2023). Each admission is converted into a standardized SOAP representation using a rule-guided conversion agent, preserving temporal evolution and diagnostic intent while nor- malizing stylistic variation. Second, 18,000 additional patient trajectories are generated using Synthea, an agent-based synthetic patient simulator designed to model longitudinal disease courses and care pathways (Walonoski et al., 2018). These synthetic records are rewrit- ten under the same SOAP specification to align structure and semantic granularity with the MIMIC- derived cohort. The combined repository thus main- tains structural uniformity while retaining clinically meaningful diversity in etiology, acuity, and dis- ease progression. Hybrid Similarity Scoring. Identifying clini- cally similar patients requires balancing strict clini- cal constraints with trajectory-level similarity, moti- vating a hybrid retrieval formulation that combines keyword alignment with semantic embeddings. To identify clinically comparable cases, we rank candidate patients using a hybrid similarity formu- lation that combines keyword-conditioned align- ment with embedding-based semantic proximity. The keyword component emphasizes discrete clini- cal signals such as diagnoses, comorbidities, and key interventions, while the semantic component captures graded similarity in longitudinal trajec- tories and response patterns. This dual formula- tion allows retrieval to satisfy eligibility-level con- straints while remaining robust to surface-level lex- ical variation. For conciseness, we present only the high-level formulation here; detailed mathematical definitions are provided in the appendix. 3.4 Unified Retrieval Workflow At inference time, a query case or question is em- bedded and used to retrieve semantically relevant guideline artifacts and patient cases from the vec- tor index. In parallel, graph-conditioned retrieval identifies guideline communities whose relational structure aligns with the query, enabling access to multi-hop clinical logic such as contraindications, escalation pathways, and applicability constraints. Retrieved guideline evidence and similar patient cases are then aggregated into a unified evidence set comprising ranked guideline communities with supporting nodes and ranked patient cases with aligned attributes. All retrieved items are presented with explicit source links, allowing users to inspect provenance and understand the basis of the sys- tem’s recommendations. Details of the GraphRAG construction and re- trieval pipeline are provided in Appendix B. 4 Evaluation and Demonstration We evaluate the proposed framework on clinical note generation and medical question answering (QA), and additionally present an interactive sys- tem demonstration. 4.1 Tasks and Setup Clinical Note Generation.We sample 1,000 de- identified ICU notes from MIMIC-IV and convert each into standardized SOAP format (Weed, 1968). The model receives the Subjective (S), Objective (O), and Assessment (A) sections and generates the corresponding Plan (P). Medical QA.We evaluate on MedQA (Jin et al., 2021) and the clinical subset of MMLU (Hendrycks et al., 2021) using the same retrieval-augmented pipeline as in the clinical setting. The similarity mixing weightλis tuned on a 100- case validation set and fixed across all experiments. 4.2 Metrics and Baselines For note generation, we report ROUGE-L and BERTScore-F 1 ; additional metrics are provided in the appendix. For QA, we report standard accuracy. We compare against strong parametric-only LLM baselines (DeepSeek-Chat, GPT-4.1-mini, and Gemini 2.5) under identical prompting and decoding settings. 4.3 Main Results Table 1 summarizes performance across tasks. Re- trieval augmentation consistently improves both generation and QA performance.Combining similar-patient retrieval and GraphRAG yields the strongest results across all backbone models. 4 Model Text GenerationQA MIMIC-IV Clinical NotesMedQAMMLU-clinical BLEUMETEORROUGE-1ROUGE-2ROUGE-LBERTScore-F1AccAcc Baselines without Retrieval DeepSeek-Chat0.0060.2510.2050.0420.1200.8090.8090.857 GPT-4.1-mini0.0120.2820.3010.0640.1440.8250.8460.868 Gemini 2.50.0140.2890.2950.0700.1470.8290.8670.875 with RAG DeepSeek-Chat0.0150.2720.2290.0570.1570.7990.8510.875 GPT-4.1-mini0.0190.2910.3510.0810.2210.8310.8850.887 Gemini 2.50.0200.3160.3680.0990.2140.8330.8920.905 Our Framework DeepSeek-Chat0.0290.2870.3400.0970.2030.8390.8640.910 GPT-4.1-mini0.0350.3210.3800.1240.2910.8510.9120.971 Gemini 2.50.0420.3340.4190.1550.3000.8620.9370.971 Table 1: Performance under different retrieval configurations on clinical note generation and QA benchmarks. 4.4 Ablation Table 2 isolates the contribution of each retrieval component.Both similar-patient retrieval and GraphRAG knowledge retrieval independently im- prove performance, with the combined framework achieving the largest gains. 4.5 Interactive System We deploy the proposed framework as a public HuggingFace Space. 1 The interface provides end- to-end interactive clinical exploration grounded in dual evidence retrieval, enabling clinicians to in- spect, compare, and reason over guideline-level and patient-level evidence within a unified decision- support workflow. Dual Evidence Control. Given a free-text clin- ical case, the system retrieves (i) similar patient records from the case repository and (i) guideline- level evidence via the GraphRAG pipeline. Users may selectively enable or disable each evidence source through explicit interface controls, allowing controllable reasoning under patient-level evidence, guideline-level knowledge, or their combination. The selected evidence is automatically packaged with the query context for downstream generation and question answering. Query-Conditioned Saliency Highlighting.To improve interpretability during evidence inspection, the similar-patient panel incorporates a lightweight query-conditioned saliency overlay. Each extracted 1 https://huggingface.co/spaces/Cryo3978/Med_ GraphRAG clinical concept (e.g., symptoms, diagnoses, comor- bidities) is assigned an importance score based on its semantic similarity to the input query embed- ding. The score is discretized into two levels and vi- sually encoded (yellow: important; red: highly important), enabling users to quickly identify case attributes most relevant to the current clinical ques- tion. Figure 2 illustrates interactive question answer- ing with controllable evidence selection.Fig- ure 3 shows the similar-patient view with query- conditioned saliency highlighting. Figure 2: Interactive question answering with dual evi- dence control. 5 Setting Text GenerationQA MIMIC-IV Clinical NotesMedQAMMLU-clinical BLEUMETEORROUGE-1ROUGE-2ROUGE-LBERTScore-F1AccAcc Baseline0.0120.2820.3010.0640.1440.8250.8460.868 + SPR0.0280.3150.3240.0990.1590.8280.8510.914 + GraphRAG0.0310.3100.3520.1150.2580.8400.8700.952 + Both0.0350.3210.3800.1240.2910.8510.9120.971 Table 2: Ablation results isolating the contribution of similar-patient retrieval (SPR) and GraphRAG-based guideline reasoning. Figure 3: Similar-patient panel with query-conditioned saliency highlighting. Clinical concepts are color-coded according to their semantic relevance to the input query (yellow: important; red: highly important). Additional screenshots and a step-by-step walk- through are provided in Appendix A. 5 Limitations The proposed system has several limitations. First, the similar-patient repository combines SOAP- normalized MIMIC-IV records with Synthea- generated synthetic cases, resulting in heteroge- neous data distributions that only partially reflect real-world clinical documentation and temporal complexity. Second, the guideline knowledge graph abstracts narrative recommendations into structured relations, which may omit contextual nu- ances present in full guideline texts. Third, the sys- tem is a research prototype intended for decision- support and evidence inspection rather than au- tonomous clinical use. Finally, evaluation is limited to note completion and medical QA benchmarks, and no gold-standard dataset currently exists for assessing patient-level similarity retrieval. Future work will explore real-world clinical de- ployment settings, expanded guideline coverage, and uncertainty-aware reasoning mechanisms to further improve safety, robustness, and practical applicability. 6 Conclusion We presented MED-COPILOT, an interactive clini- cal decision-support system designed to assist clin- icians and trainees in transparent and evidence- aware reasoning.By integrating guideline- grounded GraphRAG retrieval with similar-patient evidence within a unified and interpretable pipeline, the system enables clinicians to inspect, compare, and reason over multi-source clinical evidence. Experimental results demonstrate that combining structured guideline knowledge with patient-level analogical retrieval provides complementary ben- efits across clinical note generation and medical QA tasks, highlighting the value of controllable and inspectable evidence integration for clinical AI systems. Future work will focus on expanding guideline coverage, improving uncertainty-aware reasoning, and exploring deployment in real-world clinical workflows. Ethical Statement This work uses the MIMIC-IV database under its approved data-use agreement, and all authors com- pleted the required credentialing. Synthetic cases generated with Synthea do not correspond to real patients. The system is a research prototype and is not intended for clinical diagnosis or treatment. 6 References Daniel Alcaide and Jan Aerts. 2021. A visual analytic approach for the identification of icu patient subpop- ulations using icd diagnostic codes. PeerJ Computer Science, 7:e430. Elham Asgari, Nina Montaña-Brown, Magda Dubois, Saleh Khalil, Jasmine Balloch, Joshua Au Yeung, and Dominic Pimenta. 2025. A framework to assess clin- ical safety and hallucination rates of llms for medical text summarisation. npj Digital Medicine, 8(1):274. Andrew L Beam and Isaac S Kohane. 2018.Big data and machine learning in health care. Jama, 319(13):1317–1318. Andre Esteva, Alexandre Robicquet, Bharath Ramsun- dar, Volodymyr Kuleshov, Mark DePristo, Katherine Chou, Claire Cui, Greg Corrado, Sebastian Thrun, and Jeff Dean. 2019. A guide to deep learning in healthcare. Nature medicine, 25(1):24–29. National Institute for Health and Care Excellence. 2024. Nice clinical guidelines and knowledge summaries. Available from: https://w.nice.org.uk/guidance. Shailja Gupta, Rajesh Ranjan, and Surya Narayan Singh. 2024. A comprehensive survey of retrieval- augmented generation (rag): Evolution, current landscape and future directions.arXiv preprint arXiv:2410.12837. Dan Hendrycks and 1 others. 2021. Measuring massive multitask language understanding. ICLR. Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and 1 oth- ers. 2025. A survey on hallucination in large lan- guage models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43(2):1–55. Gautier Izacard and Edouard Grave. 2021. Leveraging passage retrieval with generative models for open do- main question answering. In Proceedings of the 16th conference of the european chapter of the association for computational linguistics: main volume, pages 874–880. Di Jin and 1 others. 2021. What disease does this patient have? a large-scale open-domain medical qa dataset. In EMNLP. Alistair EW Johnson, Lucas Bulgarelli, Lu Shen, Alvin Gayles, Ayad Shammout, Steven Horng, Tom J Pol- lard, Sicheng Hao, Benjamin Moody, Brian Gow, and 1 others. 2023. Mimic-iv, a freely accessible elec- tronic health record dataset. Scientific data, 10(1):1. Jonathan Kim, Anna Podlasek, Kie Shidara, Feng Liu, Ahmed Alaa, and Danilo Bernardo. 2025. Limita- tions of large language models in clinical problem- solving arising from inflexible reasoning. Scientific reports, 15(1):39426. Joon Lee, David M Maslove, and Joel A Dubin. 2015. Personalized mortality prediction driven by elec- tronic medical data and a patient similarity metric. PloS one, 10(5):e0127428. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Hein- rich Küttler, Mike Lewis, Wen-tau Yih, Tim Rock- täschel, and 1 others. 2020. Retrieval-augmented gen- eration for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33:9459– 9474. Lei Liang, Zhongpu Bo, Zhengke Gui, Zhongshu Zhu, Ling Zhong, Peilong Zhao, Mengshu Sun, Zhiqiang Zhang, Jun Zhou, Wenguang Chen, and 1 others. 2025. Kag: Boosting llms in professional domains via knowledge augmented generation. In Companion Proceedings of the ACM on Web Conference 2025, pages 334–343. Yuxing Lu, Xukai Zhao, and Jinzhuo Wang. 2024. Clin- icalrag: Enhancing clinical decision support through heterogeneous knowledge retrieval. In Proceedings of the 1st Workshop on Towards Knowledgeable Lan- guage Models (KnowLLM 2024), pages 64–68. Manfu Ma, Penghui Sun, Yong Li, and Weilong Huo. 2023. Predicting the risk of mortality in icu patients based on dynamic graph attention network of patient similarity. Mathematical Biosciences and Engineer- ing, 20(8):15326–15345. Zabir Al Nazi and Wei Peng. 2024. Large language models in healthcare and medical domain: A review. In Informatics, volume 11, page 57. MDPI. WorldHealthOrganization.2024.Who guidelinesrepository.Availablefrom: https://w.who.int/publications/who-guidelines. Shrey Pandit, Jiawei Xu, Junyuan Hong, Zhangyang Wang, Tianlong Chen, Kaidi Xu, and Ying Ding. 2025. Medhallu: A comprehensive benchmark for detecting medical hallucinations in large language models. arXiv preprint arXiv:2502.14302. Alvin Rajkomar, Jeffrey Dean, and Isaac Kohane. 2019. Machine learning in medicine. New England Journal of Medicine, 380(14):1347–1358. Alvin Rajkomar, Eyal Oren, Kai Chen, Andrew M Dai, Nissan Hajaj, Michaela Hardt, Peter J Liu, Xiaobing Liu, Jake Marcus, Mimi Sun, and 1 others. 2018. Scalable and accurate deep learning with electronic health records. NPJ digital medicine, 1(1):18. Diego Sanmartin. 2024. Kg-rag: Bridging the gap between knowledge and creativity. arXiv preprint arXiv:2405.12035. Tharun Sekar, Kushal, Supprethaa Shankar, Sabah Mo- hammed, and Jinan Fiaidhi. 2025. Investigations on using evidence-based graphrag pipeline using llm tailored for usmle style questions. medRxiv, pages 2025–05. 7 Anis Sharafoddini, Joel A Dubin, and Joon Lee. 2017. Patient similarity in prediction models based on health data: a scoping review. JMIR medical in- formatics, 5(1):e6730. Jiwoong Sohn, Yein Park, Chanwoong Yoon, Sihyeon Park, Hyeon Hwang, Mujeen Sung, Hyunjae Kim, and Jaewoo Kang. 2025. Rationale-guided retrieval augmented generation for medical question answer- ing. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 12739– 12753. Duolin Sun, Dan Yang, Yue Shen, Yihan Jiao, Zhehao Tan, Jie Feng, Lianzhen Zhong, Jian Wang, Peng Wei, and Jinjie Gu. 2025. Hanrag: Heuristic accu- rate noise-resistant retrieval-augmented generation for multi-hop question answering. arXiv preprint arXiv:2509.09713. Zhaohong Sun, Xudong Lu, Huilong Duan, and Haomin Li. 2022. Deep dynamic patient similarity analysis: Model development and validation in icu. Computer Methods and Programs in Biomedicine, 225:107033. Yixuan Tang and Yi Yang. 2024. Multihop-rag: Bench- marking retrieval-augmented generation for multi- hop queries. arXiv preprint arXiv:2401.15391. Jason Walonoski, Mark Kramer, Joseph Nichols, Andre Quina, Chris Moesel, Dylan Hall, Carlton Duffett, Kudakwashe Dube, Thomas Gallagher, and Scott McLachlan. 2018. Synthea: An approach, method, and software mechanism for generating synthetic pa- tients and the synthetic electronic health care record. Journal of the American Medical Informatics Associ- ation, 25(3):230–238. Lawrence Weed. 1968. Medical records, medical ed- ucation, and patient care. The Problem-Oriented Medical Record. Junde Wu, Jiayuan Zhu, Yunli Qi, Jingkun Chen, Min Xu, Filippo Menolascina, Yueming Jin, and Vicente Grau. 2025. Medical graph RAG: Evidence-based medical large language model via graph retrieval- augmented generation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 28443– 28467, Vienna, Austria. Association for Computa- tional Linguistics. Guangzhi Xiong, Qiao Jin, Zhiyong Lu, and Aidong Zhang. 2024a. Benchmarking retrieval-augmented generation for medicine. In Findings of the Associa- tion for Computational Linguistics ACL 2024, pages 6233–6251. Guangzhi Xiong, Qiao Jin, Xiao Wang, Minjia Zhang, Zhiyong Lu, and Aidong Zhang. 2024b. Improving retrieval-augmented generation in medicine with it- erative follow-up questions. In Biocomputing 2025: Proceedings of the Pacific Symposium, pages 199– 214. World Scientific. Rui Yang, Haoran Liu, Edison Marrese-Taylor, Qingcheng Zeng, Yuhe Ke, Wanxin Li, Lechao Cheng, Qingyu Chen, James Caverlee, Yutaka Mat- suo, and 1 others. 2024. Kg-rank: Enhancing large language models for medical qa with knowledge graphs and ranking techniques. In Proceedings of the 23rd Workshop on Biomedical Natural Language Processing, pages 155–166. Michihiro Yasunaga, Hongyu Ren, Antoine Bosse- lut, Percy Liang, and Jure Leskovec. 2021. Qa- gnn: Reasoning with language models and knowl- edge graphs for question answering. arXiv preprint arXiv:2104.06378. Qinggang Zhang, Shengyuan Chen, Yuanchen Bei, Zheng Yuan, Huachi Zhou, Zijin Hong, Hao Chen, Yilin Xiao, Chuang Zhou, Junnan Dong, and 1 others. 2025. A survey of graph retrieval-augmented gener- ation for customized large language models. arXiv preprint arXiv:2501.13958. Zhengyun Zhao, Qiao Jin, Fangyuan Chen, Tuorui Peng, and Sheng Yu. 2023. A large-scale dataset of patient summaries for retrieval-based clinical decision sup- port systems. Scientific data, 10(1):909. 8 A Demo Walkthrough This appendix provides a step-by-step walkthrough of the interactive demo interface. A.1 Case Input and Query Locking Figure 4: Patient case input and query locking interface. A.2 Similar Patient Retrieval Figure 5: Retrieval of the most similar patient case from the curated repository. Following case locking, the system retrieves clini- cally similar patient records from the 36,000-case repository using the proposed hybrid similarity scoring function. Figure 5 shows the most sim- ilar patient identified for the input case. Clini- cally salient attributes, including symptoms, co- morbidities, diagnostic findings, and interventions, are highlighted to support rapid comparison and interpretability. This retrieved case provides experi- ential, case-based context grounded in prior patient trajectories. A.3 Guideline Retrieval via GraphRAG Figure 6:Guideline-grounded evidence retrieval through the GraphRAG pipeline. In parallel with similar-patient retrieval, the system performs guideline-grounded retrieval using the GraphRAG pipeline. As illustrated in Figure 6, rel- evant clinical recommendations are identified and 9 summarized from authoritative WHO and NICE guidelines. Retrieved guideline content exposes structured evidence such as treatment options, con- traindications, and escalation criteria. A.4 Interactive Evidence-Aware Question Answering Figure 7: Interactive question answering with user- controlled evidence selection. Figure 7 demonstrates the interactive question- answering interface. Users may selectively choose whether to incorporate similar patient cases, guide- line summaries, or both into downstream reason- ing. Based on the selected evidence sources, the system dynamically constructs a structured prompt combining the locked patient case and retrieved ev- idence, enabling transparent and controllable clin- ical exploration. Generated responses explicitly reflect the chosen evidence context, allowing users to examine how different information sources in- fluence the model’s reasoning. B Appendix B: GraphRAG Construction Details Figure 8: Detailed construction and retrieval pipeline of the GraphRAG module, including segmentation, graph extraction, indexing, and hybrid retrieval. 10