DiaQwen: A Knowledge-Enhanced LLM Framework for Privacy-Preserving Diabetes Diagnosis

Diabetes remains a leading chronic disease worldwide, yet applying large language models (LLMs) to its diagnosis faces four barriers: insufficient domain knowledge, limited multi-turn reasoning, inability to deploy locally under privacy constraints, and high hallucination rates. We present DiaQwen, a privacy-preserving diagnostic system built on Qwen-4B that runs on a single consumer GPU. DiaQwen integrates three innovations: (1)KBase, a diabetes knowledge base encoding over 2000 clinical triplets from authoritative guidelines, EHRs, and public QA datasets; (2)HyTrain, a hybrid fine-tuning strategy combining LoRA with epoch-level self-distillation; and (3)DyQA, a dynamic consultation module navigating a diagnostic graph via A* search. On 2000 test cases, DiaQwen achieves 93.4% accuracy with 0.7% hallucination rate, surpassing baselines by over 4 points while producing guideline-aligned explanations (BLEU=0.72, ROUGE-L=0.96).