Diabetes AI combines knowledge graphs with language models for traceable answers

A research team has built a diabetes question-answering system that links a structured medical knowledge graph to a retrieval-augmented large language model, allowing the software to display the graph paths and guideline passages used to construct each response. The approach was tested primarily on question-parsing components rather than on the clinical correctness of complete answers, so the results show technical feasibility but do not establish that the system is ready to advise patients.

The prototype is designed around a central problem in generative artificial intelligence: language models can produce fluent medical explanations without making the information behind them easy to inspect. The researchers say their architecture addresses this by keeping two evidence streams separate. One stream comes from DiaKG, a diabetes knowledge graph built from Chinese expert consensus and authoritative guidance; the other comes from passages retrieved from a national primary-care diabetes guideline. The final model, GLM-4-Flash, receives both streams in a structured prompt, while an interface exposes recognized entities, predicted intent, graph queries, relation paths and retrieved paragraphs.

DiaKG was assembled from the Ruijin Hospital MMC dataset for AI-assisted diabetes knowledge-graph construction on Alibaba’s Tianchi platform. It contains nearly 190,000 annotated Chinese medical entities, including drugs, symptoms and diseases, along with 6,890 annotated relations and draws on almost a decade of Chinese expert consensus and authoritative diabetes guidelines. The researchers imported the annotations into the Neo4j graph database, yielding more than 4,900 distinct entity nodes, 18 entity types and 15 relation types. Relations concentrate mainly on diseases and drugs, including links involving treatment, adverse drug events, dosage, duration and frequency.

When a user asks a question, the system first parses it in two stages. Qwen2.5-1.5B, fine-tuned with low-rank adaptation, recognizes medical entities, while a hybrid BERT-CNN classifier predicts the question’s intent. The intent is then translated through predefined rules into one or more graph relations, allowing the software to construct Cypher queries. The graph engine supports expansion around a single entity and relation discovery between multiple entities, including direct connections and indirect paths of up to three hops. A breadth-first search was chosen because preliminary tests indicated that longer paths were often less medically relevant and substantially slower to retrieve.

The entity-recognition experiments produced weighted F1 scores of 71.34% across the tested data, with the best dynamic learning-rate schedule outperforming a fixed rate by 2.4 percentage points. That schedule combined a linear warmup during the first 10% of training steps with cosine decay afterward, while L2 regularization was used to limit overfitting. Strict entity-boundary evaluation gave an F1 of 81.02% for diseases and 85.27% for drugs. Only about 1.38 million low-rank parameters were trained, roughly 0.092% of the base Qwen model’s parameters, demonstrating a parameter-efficient route to adapting the model for a relatively small specialist dataset.

Drug names may be easier to identify because many follow recognizable naming patterns, whereas disease mentions can overlap with symptoms or more specific diagnoses. The authors also note that drug entities are more numerous in the annotations, giving the model more examples from which to learn stable boundaries. These are dataset-level observations, however, and do not establish how the parser would perform on unfamiliar names, informal descriptions or languages other than Chinese. The reported cross-validation results are component measurements rather than evidence of end-to-end question-answering accuracy.

For intent classification, the researchers trained and compared BiLSTM, TextCNN, an attention-enhanced BiLSTM and several pretrained language models on the DaCorp Chinese diabetes question corpus. The hybrid BERT-CNN architecture achieved 84.31% accuracy, 75.08% recall and an F1 score of 73.38%, outperforming the reported baselines on those metrics. BERT alone reached 83.5% accuracy, while the hybrid added 0.81 percentage points in accuracy and approximately 1.4 points in F1. Its design combines BERT’s contextual representation of a complete question with a one-dimensional convolutional layer designed to detect short phrase patterns that may signal particular intents, such as asking about causes, prevention or treatment.

Guideline retrieval adds a second evidence channel. The researchers extracted and segmented the “National Primary Care Diabetes Prevention and Treatment Guideline (2022),” then encoded 189,215 passages with the Chinese-optimized BCE embedding model and stored them in a vector database. By default, the implemented system returns the three passages most similar to the question. These passages are not merged invisibly with graph facts: the prompt presents knowledge-graph evidence and retrieved guideline evidence as distinct fields and instructs the language model to remain cautious when evidence is incomplete. LangChain coordinates the components, while a Streamlit interface displays streaming answers and expandable analysis details.

Public code and three intent-classification checkpoints have been released, allowing researchers to inspect or reproduce the graph construction, query, retrieval and training components. The materials do not include an expert-annotated benchmark of complete medical answers. The researchers therefore call for end-to-end testing by qualified specialists, with measures including clinical correctness, completeness, consistency with cited guidance, hallucination rate, interpretability and agreement between reviewers. Controlled comparisons among a language-model-only system, graph-only retrieval, guideline retrieval alone and the combined architecture would also be needed to determine how much each evidence source contributes.

The researchers also identify unresolved deployment questions involving multilingual, colloquial, code-mixed, emotionally charged and urgent queries. They say patient-facing use would require prospective clinical validation, privacy protection, risk management and clinician supervision, along with measurements of latency, concurrent use and deployment cost. The study’s most concrete achievement is therefore an inspectable pipeline: its parsers performed strongly on the evaluated component tasks, and its architecture makes upstream evidence visible. Whether that visibility produces reliably faithful medical answers remains an open question that requires independent, expert-rated validation.

Subject of Research: Evaluation of a traceable diabetes question-answering framework integrating a knowledge graph, retrieval-augmented generation and large language models.

Article Title: Diabetes question answering method based on knowledge graph and large language model

Article References: Zhang, Y., Li, S., Lyu, T., Yu, X., Jiang, T., Li, X., He, G., & Zhou, W. (2026). Diabetes question answering method based on knowledge graph and large language model. Discover Artificial Intelligence, 6(1), Article 1206. https://doi.org/10.1007/s44163-026-02078-2

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02078-2

Keywords: diabetes, medical AI, question answering, knowledge graphs, retrieval-augmented generation, large language models, entity recognition, intent classification, explainable AI, clinical validation, question, answering