This retrospective study used encrypted and deidentified data from West China Hospital involving no patient privacy–sensitive information or clinical interventions. Data extraction and study procedures were approved by the ethics committee of West China Hospital, Sichuan University (approval 2024-126), in accordance with informed consent requirements.

This study used the big data integration platform of West China Hospital as the primary source of data [10]. We collected electronic health record data from patients diagnosed with diabetes who visited the outpatient department from January 2022 to February 2022. The collected data included information such as the patients’ department of visit, age, gender, chief concern, present illness history, and diagnosis, which served as input for the model. In addition, we obtained data on patients’ outpatient medications, laboratory test items, examinations, and physician opinions, which were used as model outputs. Patient data with missing chief concerns and present illness history were excluded to ensure data quality and usability. Furthermore, we collected an additional set of data from 300 patients (visiting in March 2022) as a test set to evaluate the performance of the optimized model.

Before model training, we performed systematic preprocessing on the collected patient data, including data cleaning, standardization, and anonymization. Data cleaning involved removing missing values, outliers, and duplicate records to ensure the integrity and reliability of the model inputs. Standardization ensured consistency in format, units, and value ranges across different data types. For text data, including chief concerns, present illness history, examinations, and laboratory test results, we applied cleaning, terminology unification, and segmenting of key information while removing sensitive information to protect patient privacy. The processed data were then organized into a question-answer format (instruction-input-output) suitable for LLM fine-tuning: patient basic information and clinical text (eg, department, gender, age, diagnosis, chief concern, and present illness history) were used as the input, and corresponding medical orders, examinations, recommended laboratory tests, and treatment recommendations were integrated as the output (shown in Textbox 1). This approach ensured that both structured and unstructured information was effectively used for model training. The dataset was then split into training and test sets in an 8:2 ratio.

In this study, we compared 3 LLMs for clinical text prediction tasks: Llama 3.1-8B (Meta AI) [11], Qwen3-8B (Alibaba Cloud) [12], and GLM4-9B (THUDM/Z.ai) [13]. Llama 3.1-8B is an open-source model optimized for efficient generalization and complex text understanding. Qwen3-8B is a multilingual LLM with strong capabilities in both structured and unstructured medical text processing. GLM4-9B, developed based on the general language model architecture, is designed for bilingual question-answering tasks and supports local deployment through model quantization. All models were fine-tuned and evaluated using a single H100 graphics processing unit (NVIDIA). Smaller model sizes were chosen to facilitate clinical deployment and wider adoption in practical settings while still maintaining competitive performance for downstream tasks.

To adapt the selected LLMs for clinical treatment recommendation tasks, we used a parameter-efficient fine-tuning strategy that combines instruction-based prompting with low-rank adaptation (LoRA) [14]. Instruction templates were designed to explicitly guide the model in interpreting patient information and generating clinically appropriate treatment suggestions. Building on this, LoRA was applied to the attention layers to enable efficient task-specific adaptation while updating only a small number of parameters. During fine-tuning, we explored multiple hyperparameter configurations, including learning rate, batch size, and LoRA-specific parameters such as rank and scaling factor. Model performance was assessed on a held-out validation set using both automatic text generation metrics (eg, Bilingual Evaluation Understudy [15] and Recall-Oriented Understudy for Gisting Evaluation [16]) and clinically oriented evaluation criteria, including the correctness and appropriateness of the recommended treatments. This combined approach provided a balanced trade-off between computational efficiency and clinical relevance, offering practical guidance for deploying LLM-based treatment recommendation systems in real-world clinical settings.

To ensure a rigorous and clinically meaningful evaluation of our LLM-generated diabetes treatment recommendations, we implemented a structured, multidimensional physician assessment protocol. Six dimensions were assessed: treatment appropriateness, medication accuracy, relevance of suggested examinations, safety, logical reasoning, and overall clinical usefulness. Each dimension was rated using a standardized 5-point Likert scale [17] ranging from 1 (“completely unreasonable or not useful”) to 5 (“fully reasonable and clinically valuable”). Five board-certified endocrinologists (each with more than 5 years of independent clinical practice) served as expert raters. All assessments were performed independently and in a fully blinded manner: raters were unaware of which LLM produced each recommendation and were prohibited from discussing cases. Each physician reviewed the same 300 treatment recommendations generated by the 3 LLMs, including both base and fine-tuned versions. Recommendations included medication plans, dosage adjustments, and suggested laboratory tests. This protocol ensured a systematic, reproducible, and clinician-centered evaluation of model outputs, enabling identification of potential risks and areas requiring refinement.

We implemented a retrieval-augmented generation (RAG) [18] framework to integrate hospital knowledge resources, including the medical order database, diabetes treatment guidelines, and clinical protocols, enabling the model to dynamically retrieve relevant knowledge during treatment recommendations and provide context-specific, up-to-date clinical information. An agent system was developed to interface with the hospital information system. It handles data cleaning, integration, and monitoring by extracting and standardizing patient data, consolidating heterogeneous sources into a structured format suitable for RAG, and ensuring data integrity and consistency. By combining RAG with agent-driven data management, the system efficiently leverages internal knowledge bases to support clinically grounded, accurate, and interpretable treatment recommendations. The overall workflow of this clinical application system is illustrated in Figure 1.

As shown in Figure 2, the final dataset comprised 20,619 patients, with 80% allocated for model training and 20% held out as a test set for performance evaluation. Figure 2 illustrates the data collection process.

All 3 LLMs—Llama 3.1-8B, Qwen3-8B, and GLM4-9B—were evaluated before and after LoRA fine-tuning. Their training and test loss curves are shown in Figure 3, and model performance was assessed using both automatic text generation metrics and clinical physician evaluations.

Table 1 summarizes the comparative performance of the baseline and fine-tuned models measured using Bilingual Evaluation Understudy for 4-grams (BLEU-4), Recall-Oriented Understudy for Gisting Evaluation for overlap of unigrams (ROUGE-1), Recall-Oriented Understudy for Gisting Evaluation for overlap of bigrams (ROUGE-2), and Recall-Oriented Understudy for Gisting Evaluation–Longest Common Subsequence (ROUGE-L). Before fine-tuning, the 3 base models demonstrated moderate capability in generating clinically relevant recommendations, with mean BLEU-4 scores ranging from 45.13 (SD 1.98) to 50.84 (SD 1.87) and ROUGE-L scores ranging from 9.83  (SD 0.54) to 13.90 (SD 0.74). Among the base models, GLM4-9B achieved the highest performance across all metrics. After fine-tuning, all models showed significant gains in both lexical similarity and content relevance. The BLEU-4 mean score increased by 14.48 points for Llama 3.1-8B, 15.95 points for Qwen3-8B, and 17.09 points for GLM4-9B. ROUGE-1 and ROUGE-L exhibited similar patterns, with improvements exceeding 20 points for all 3 models. The fine-tuned GLM4-9B outperformed all models, achieving the highest BLEU-4 (mean 67.93, SD 2.74), ROUGE-1 (mean 44.30, SD 3.91), ROUGE-2 (mean 27.34, SD 1.85), and ROUGE-L (mean 37.67, SD 2.88) scores.

Regarding the clinical evaluation, the performance of the original models and their LoRA fine-tuned versions is summarized in Textbox 2.

Table 2 shows the results of 5 endocrinologists’ evaluation of the 300 treatment recommendations generated by the 3 fine-tuned LLMs. Overall, the mean scores indicated that most recommendations were clinically relevant to diabetes management (mean usefulness scores above 4 for all fine-tuned models), whereas the base models received substantially lower ratings. Approximately 10% of the recommendations (32/300, 10.7%) were judged by at least one rater as potentially risky (eg, inappropriate escalation or overtreatment), underscoring the need for human oversight.

As shown in Figure 4, the radar chart highlights the multidimensional improvements achieved through LoRA fine-tuning across all 6 evaluation metrics. For all 3 models, the fine-tuned versions consistently exhibited higher scores than their base counterparts. Among the models, GLM4-9B (fine-tuned) achieved the highest performance across nearly all dimensions, particularly in treatment appropriateness, relevance, and overall usefulness.

This study systematically evaluated the performance of 3 LLMs (Llama 3.1-8B, Qwen3-8B, and GLM4-9B) on a clinical treatment recommendation task for diabetes and explored a parameter-efficient optimization strategy combining instruction-based prompting and LoRA fine-tuning. It should be noted that all 3 models are general-purpose open-source LLMs and were not specifically fine-tuned on medical data, serving as preliminary baselines for clinical recommendation tasks. The untuned base models achieved moderate performance in generating clinically relevant recommendations, with mean BLEU-4 scores ranging from 45.13 to 50.84 and mean ROUGE-L scores ranging from 9.83 to 13.90, among which the base GLM4-9B model performed best across metrics. These findings indicate that even relatively small models possess nontrivial text generation capabilities but exhibit limitations in content relevance and accuracy for specialized clinical tasks. After LoRA fine-tuning, all models showed substantial improvements on both automatic evaluation metrics and clinician assessments. BLEU-4 scores increased by approximately 14 to 17 absolute points on average, whereas ROUGE-1 and ROUGE-L scores improved by more than 20 points, suggesting that fine-tuning effectively enhanced lexical similarity and information completeness in the generated texts. In clinician evaluations, scores increased markedly across the 6 dimensions—therapeutic appropriateness, medication accuracy, examination relevance, safety, logical reasoning, and overall usefulness. Among them, fine-tuned GLM4-9B achieved the best performance on key dimensions such as therapeutic appropriateness, relevance, and overall usefulness, whereas fine-tuned Llama 3.1-8B and Qwen3-8B also demonstrated varying degrees of improvement (Figure 4). The radar chart clearly illustrates these multidimensional gains, underscoring the effectiveness of combining LoRA with instruction-based prompting to enhance clinical applicability. Furthermore, by leveraging an RAG framework and an agent-based data processing and knowledge retrieval pipeline (Figure 1), the models can dynamically access up-to-date institutional clinical guidelines, medication order databases, and diabetes treatment standards, thereby enabling individualized, context-aware clinical recommendations. This approach not only improves the accuracy and safety of the generated suggestions but also strengthens their interpretability and practical usability in clinical settings. After repeated debugging and testing of the model, we found that, for a subset of patients with complex medical records, the model output can be potentially harmful and fail to assist health care professionals in their treatment. This situation can occur for the following reasons:

This study has several limitations. First, the proposed framework has been evaluated only in a controlled experimental setting and has not yet been prospectively validated in real-world clinical workflows. As such, its clinical effectiveness and operational impact remain to be systematically assessed. Second, we only used relatively small-scale models, which have substantially fewer parameters than larger frontier models such as DeepSeek R1 [22], Qwen3-32B, or Llama 3-65B. While the smaller models reduce computational requirements and facilitate low-cost intranet deployment, they may limit performance and generalization compared with larger models. Third, the training data for this study were collected from a single medical institution and are relatively limited in size, which may affect the robustness and external validity of the results. It is important to understand that, while LLMs often perform well in many scenarios, they may have limitations when dealing with complex medical records. Therefore, model outputs should not be the sole basis for decision-making. Health care professionals should rely on their expertise and clinical judgment and integrate model outputs with comprehensive assessments to make informed decisions [23]. Additionally, we plan to fine-tune domain-specific medical LLMs (such as MedGemma 27B [24]) and explore the integration of medical knowledge graphs to incorporate structured clinical knowledge, enhance model reasoning [25], and further improve the consistency and reliability of the generated recommendations. Future research will also focus on leveraging larger models and multi-institutional datasets to strengthen model performance, generalizability, and robustness.

Overall, this study validates the feasibility and effectiveness of combining small-scale LLMs with LoRA fine-tuning and an RAG- and agent-assisted data processing and knowledge retrieval strategy for clinical treatment recommendations in diabetes. The fine-tuned models not only achieved superior performance on automated text generation metrics but also generated treatment recommendations deemed safe, clinically appropriate, and of substantial reference value in structured clinician evaluations. These findings provide a viable pathway for the responsible deployment of LLMs in real-world medical applications. Future research should scale up training datasets; extend validation to a broader range of disease entities; and incorporate longitudinal real-world evidence to further assess long-term clinical effectiveness, safety, and generalizability.