As diabetes is characterized by youthfulness and an increase of diabetes in obese and overweight populations, diabetes prevention and treatment (DPT) becomes important. Recently, Large Language Models (LLM) make progress in health but do not focus on the diabetes prevention and control. Therefore, there are lack of knowledges, reasoning paths and numerical reasoning. To address these issues, a reasoning framework of ‘screening-target-intervention’ was proposed towards DPT based on LLM via personal health data. Based on authoritative guidelines and expert systems, health data from medical check-ups were integrated to construct a DPT dataset. The DPT-LLM was fine-tuned as the core of the reasoning framework by combining text serialization and structured reasoning strategy for the effective generation. Experiments on the DPT dataset show that DPT-LLM was able to recognize valid factors from numerous risk factors for accurate numerical inference, it outperforms other baselines, and demonstrates the effectiveness of the framework in diabetes prevention and treatment.