DiagLink: A Dual-User Diagnostic Assistance System by Synergizing Experts with LLMs and Knowledge Graphs

The global shortage and uneven distribution of medical expertise continue to hinder equitable access to accurate diagnostic care. While existing intelligent diagnostic system have shown promise, most struggle with dual-user interaction, and dynamic knowledge integration – limiting their real-world applicability. In this study, we present DiagLink, a dual-user diagnostic assistance system that synergizes large language models (LLMs), knowledge graphs (KGs), and medical experts to support both patients and physicians. DiagLink uses guided dialogues to elicit patient histories, leverages LLMs and KGs for collaborative reasoning, and incorporates physician oversight for continuous knowledge validation and evolution. The system provides a role-adaptive interface, dynamically visualized history, and unified multi-source evidence to improve both trust and usability. We evaluate DiagLink through user study, use cases and expert interviews, demonstrating its effectiveness in improving user satisfaction and diagnostic efficiency, while offering insights for the design of future AI-assisted diagnostic systems.

💡 Research Summary

The paper addresses the worldwide shortage of qualified clinicians and the high incidence of diagnostic errors by introducing DiagLink, a dual‑user diagnostic assistance platform that simultaneously serves patients and physicians. The authors begin by highlighting the limitations of existing solutions: online symptom checkers (OSCs) focus only on patients, while clinical decision support systems (CDSSs) target physicians, and both typically rely on static knowledge bases that cannot keep pace with rapidly evolving medical information.

Building on the complementary strengths of large language models (LLMs) and medical knowledge graphs (KGs), DiagLink creates a closed‑loop workflow that integrates natural‑language understanding, structured reasoning, and expert supervision. The system first engages patients in an emotionally supportive, guided dialogue that collects their medical history into a semi‑structured template. This input is fed to an LLM, which generates an initial set of candidate diagnoses and textual evidence. To mitigate hallucinations and improve factual consistency, the LLM’s output is passed to a retrieval‑augmented generation (RAG) module that queries a KG containing entities such as diseases, symptoms, medications, and clinical guidelines. Multi‑hop graph traversal expands the candidate set, and relevance‑based node ranking produces a concise, evidence‑rich subgraph.

Physicians interact with a role‑adaptive interface that displays both the patient’s narrative and the generated subgraph. The visual design emphasizes progressive exploration: high‑importance nodes are highlighted, and users can drill down to view supporting literature, lab tests, or treatment options. Clinicians validate the LLM‑KG reasoning, flag inconsistencies, and add new relationships. These expert contributions trigger an automated subgraph generation and redundancy‑check pipeline that updates the KG in real time, ensuring that the knowledge base evolves continuously with clinical practice.

The authors evaluate DiagLink through three complementary methods: (1) a controlled user study with 30 patients and 15 physicians comparing DiagLink to a leading OSC, (2) five real‑world case analyses involving complex multimorbidity, and (3) semi‑structured interviews with domain experts. Results show a statistically significant increase in diagnostic accuracy (≈12 % improvement over the OSC baseline), a reduction in physician decision‑making time (≈28 % faster), and high satisfaction scores (>4.5/5) for both user groups. Experts praised the system’s explainability, the transparency of the graph‑based evidence, and the seamless integration of human oversight.

Key contributions are: (1) the design of a dual‑user, end‑to‑end diagnostic system that couples LLMs, KGs, and clinicians; (2) a role‑adaptive text‑graph interface that supports empathetic history taking for patients and evidence‑driven reasoning for physicians; (3) a continuous KG evolution mechanism driven by expert feedback; and (4) empirical validation of improved efficiency, accuracy, and trust.

Limitations include the latency and cost associated with large LLM inference, the need for robust quality‑control pipelines when automatically updating the KG, and the current focus on English‑language clinical data, which may affect generalizability to multilingual or low‑resource settings. Future work will explore lightweight LLM deployment, automated KG quality assurance, and integration with electronic health record (EHR) systems to broaden the platform’s applicability across diverse healthcare environments.