💬 NLP

GraphMed-LT: Patient-Specific Graph Memory with Latent Clinical Thought Refinement for Multi-Turn Medical Conversations

GraphMed-LT is a novel framework for multi-turn medical conversations that constructs an incrementally updated, patient-specific graph memory from clinical triplets and refines it through a trainable doctor agent's hidden-state feedback, thereby overcoming fragmented evidence limitations and significantly improving diagnostic accuracy across various medical specialties.

Original authors: Zhaohan Meng, Zaiqiao Meng, Siwei Liu, Hao Xu, Ke Yuan, Iadh Ounis

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhaohan Meng, Zaiqiao Meng, Siwei Liu, Hao Xu, Ke Yuan, Iadh Ounis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the real world, a medical diagnosis is rarely a single moment of revelation. It is a conversation, a slow unfolding where a doctor asks a question, listens to the answer, and then uses that new piece of information to ask the next, more precise question. This back-and-forth is how human doctors build a picture of a patient's health, gathering clues over time rather than receiving a complete file all at once. For computers, however, simulating this process has been a significant hurdle. Traditional artificial intelligence systems designed for medical questions often treat the entire conversation as a long, unstructured list of text. As the dialogue grows, the computer struggles to keep track of which symptom belongs to which part of the story, or how a medication mentioned in the first minute connects to a pain described ten minutes later. The information is there, but it is fragmented, making it difficult for the machine to reason through the complex chain of cause and effect required for a correct diagnosis.

Researchers at the University of Glasgow, working alongside colleagues from Harvard Medical School and other institutions, have developed a new approach to solve this problem. They created a system called GraphMed-LT, which changes how an artificial intelligence doctor remembers and processes a patient's story. Instead of simply reading through a long transcript of the conversation, the system breaks the patient's words down into small, structured facts—like "fatigue" linked to "night sweats"—and organizes these facts into a living map. This map, or graph, grows and updates with every new question and answer, connecting new clues to old ones in a way that mimics how a human mind links related ideas. The system then uses this organized map to refine its own internal thinking before it decides what to ask next or what the final diagnosis should be.

The core idea behind this work is that the way information is stored matters just as much as the information itself. In their experiments, the researchers tested this system against several other advanced medical AI models using three different sets of simulated patient conversations. These conversations covered a wide range of medical specialties, from dermatology to cardiology, and involved doctors asking follow-up questions to uncover missing details. The results showed that the new system was significantly more accurate than the previous best methods. In one specific test using a powerful language model, the new system improved the accuracy of the final diagnosis by up to 6.3 percentage points compared to the strongest existing competitor. This might sound like a small number, but in the context of medical testing, it represents a substantial leap in reliability.

What makes this improvement possible is the system's ability to treat the conversation as a structured network of evidence rather than a stream of words. When a patient says they are tired and have night sweats, the system extracts these as specific data points and links them together. If the patient later mentions they are taking a specific medication, the system immediately connects that drug to the symptoms, checking a vast database of medical knowledge to see if the drug could be the cause. This happens in real-time, with the system constantly updating its internal map. Crucially, the system does not just store this map; it uses it to "think" before it speaks. It runs a refinement process where it checks its own reasoning against the organized facts, ensuring that its next question is based on a solid understanding of the entire clinical picture, not just the most recent sentence.

The researchers also found that this method led to better conversations, not just better answers. The system asked questions that were more likely to be answered by the patient, avoiding queries that were irrelevant or impossible to answer based on the available information. This suggests that the system was not just guessing or asking random questions to fill time, but was genuinely using the accumulated evidence to guide the conversation toward a solution. The study showed that this approach worked consistently well across different types of medical conditions and different underlying computer models, indicating that the method is robust and not dependent on a single specific technology.

While the study was conducted using simulated patients and controlled datasets, the results offer a clear path forward for how artificial intelligence might handle complex, multi-step tasks in the future. The work suggests that for AI to be truly effective in fields like medicine, where context and connection are everything, it must move beyond simple text processing. It needs a way to organize facts into a coherent structure that can be updated and refined as new information arrives. By giving the AI a "memory" that is structured like a map of relationships rather than a pile of documents, the researchers have shown that machines can learn to reason through a diagnosis with a level of consistency and accuracy that was previously out of reach. This does not mean the system replaces human doctors, but rather that it provides a tool that could help support clinical reasoning, ensuring that no clue is left unconnected and no question is asked without purpose.

Technical Summary: GraphMed-LT

Problem Definition

Traditional medical question answering (QA) is typically formulated as a single-turn task where all patient information is provided simultaneously. However, realistic clinical diagnosis occurs through multi-turn conversations where a doctor progressively gathers information through iterative questioning. Existing multi-turn medical conversation systems often rely on accumulated dialogue history as memory. This approach leaves clinically relevant evidence fragmented across turns, making effective multi-hop reasoning difficult. The core challenge is that clinical facts acquired in separate turns remain unstructured, limiting the system's ability to integrate evidence for reliable diagnosis.

Methodology

The authors propose GraphMed-LT, a framework designed to organize and refine patient evidence across conversation turns using a patient-specific graph memory and latent clinical thought refinement. The system operates through four main components:

1. The Patient Agent

Simulates a patient interacting with a doctor. It is constrained to answer only using information contained in a complete patient record ( $P$ ). If a query cannot be answered by the record, it returns a fixed refusal message. This ensures that subsequent evidence extraction is grounded in dataset-supported facts rather than speculation.

2. The Triplet Agent

Converts unstructured patient responses into structured data.

Extraction: Extracts patient-specific clinical triplets (entity-relation-entity) explicitly expressed in the patient's response.
Retrieval: Retrieves relevant knowledge triplets from PrimeKG (a broad biomedical knowledge graph) using similarity matching via the BiCA encoder. The top-3 relevant triplets serve as auxiliary background knowledge.

3. Patient-Specific Graph Memory

This module incrementally organizes triplets into a dynamic graph structure ( $G_t$ ).

Construction: Initialized with triplets from the initial patient description. At each turn, new clinical entities and relation-labelled edges are added. Source-aware edge types distinguish between patient-observed triplets and retrieved knowledge.
Encoding: A Graph Attention Network (GAT) encodes the graph, aggregating node representations into a graph-level representation ( $g_t$ ).
Projection: A trainable projector maps $g_t$ into the input embedding space of the doctor agent, producing a set of graph-conditioned evidence tokens ( $Z_t = [z_1, \dots, z_m]$ ).

4. Doctor Agent with Latent Clinical Thought Refinement

The doctor agent is a trainable Large Language Model (LLM) responsible for asking follow-up questions or predicting the final diagnosis.

Refinement Mechanism: Instead of treating graph evidence as a static prefix, GraphMed-LT introduces latent clinical thought refinement. Each evidence token $z_i$ is paired with a mutable thought token $h_i$ .
Iterative Process: Over $K$ refinement steps, the doctor agent processes an interleaved sequence $[z_1, h^k_1, \dots, z_m, h^k_m]$ . The hidden states at the thought-token positions are used to update the thought tokens for the next step, while the evidence tokens ( $z_i$ ) remain fixed as anchors.
Output: The final refined context is prepended to the doctor's prompt to generate the next question or final answer.

Training: The system is trained using cross-entropy loss on the final answer prediction. Only the graph encoder, projector, and doctor agent parameters are trainable; the patient and triplet agents remain frozen.

Key Contributions

GraphMed-LT Framework: A novel approach that incrementally organizes extracted patient triplets and retrieved knowledge triplets into a patient-specific graph memory for multi-turn clinical evidence modeling.
Latent Clinical Thought Refinement: A continuous-space reasoning mechanism that refines graph-conditioned evidence tokens inside a trainable doctor agent through hidden-state feedback, allowing the agent to internalize graph evidence before generating outputs.
Empirical Validation: Extensive experiments demonstrating consistent improvements over strong baselines across multiple LLM backbones and medical specialties.

Experimental Results

The authors evaluated GraphMed-LT on three multi-turn medical QA benchmarks: iCRAFT-MD, iMedQA, and iMedMCQA.

Diagnostic Accuracy: GraphMed-LT consistently outperformed seven strong baselines (including UoT, CoD, MEDIQ, and ProMed) across multiple LLM backbones (LLaMA-3-8B/70B, Qwen-2.5-7B/72B). It achieved absolute improvements of 3.3 to 6.3 percentage points over the strongest baselines. For instance, on iMedQA with Qwen-2.5-72B, it reached 73.0% accuracy compared to 67.5% for the next best baseline (ProMed).
Question Answerability (AQR): The system asked follow-up questions that were more frequently answerable under the simulator protocol (e.g., +5.1% AQR on iCRAFT-MD), indicating improved information-seeking efficiency rather than indiscriminate questioning.
Ablation Studies:
- Removing latent thought refinement caused the largest accuracy drop (e.g., from 73.0% to 66.8% on iMedQA), highlighting its centrality to internalizing evidence.
- Removing graph memory resulted in the second-largest drop, confirming the value of structured cross-turn organization.
- Removing retrieved triplets had a smaller impact, suggesting the primary gains come from the graph structure and refinement rather than external knowledge alone.
Efficiency: GraphMed-LT achieved higher accuracy than baselines while using a comparable number of interaction turns, improving the accuracy-efficiency trade-off.
Robustness: The system showed consistent gains across eight different medical specialties and remained robust to moderate triplet missingness (up to 20% dropout).

Significance and Claims

The paper claims that GraphMed-LT addresses the fragmentation of clinical evidence in multi-turn conversations by moving beyond simple dialogue history to a structured, patient-specific graph memory. The integration of latent clinical thought refinement allows the doctor agent to iteratively reconcile graph-derived evidence with the evolving clinical context before generating a response.

The authors assert that this combination enables more reliable multi-hop reasoning, leading to:

Higher diagnostic accuracy across diverse medical domains.
More effective information-seeking behaviors (asking more answerable questions).
A mechanism that supports robust multi-turn medical reasoning without relying on excessively long conversations.

The work positions itself as a step toward more realistic clinical simulation where diagnoses are formed through iterative evidence accumulation and structured reasoning, rather than single-shot processing of static data.