关键词: 终末期肾病, 维持性血液透析, 辅助决策模型, 智能化透析模式, 大语言模型

Abstract: Objective To develop an intelligent auxiliary decision-making system for hemodialysis mode based on the Qwen2.5～7B-Instruct large language model, and to evaluate its clinical application effects.  Methods Blood pressure standards and 11 dialysis mode determination rules were integrated. A structural input paradigm was established using 33 virtual cases to optimize the prompt template. Data from 192 patients with maintenance hemodialysis (MHD) at Wenzhou People's Hospital, comprising 940 treatment cycles, were used as the training set. The Low-Rank Adaptation (LoRA) technique was employed for parameter fine-tuning and multi-round prompt engineering optimization. The model performance was evaluated using Kappa consistency, accuracy, precision, recall, and F1 score. A prospective validation study was conducted with 232 MHD patients from April to May 2025. They were divided into the model group (n=58) and the control group (n=174). Blood pressure control, incidence of complications, and evaluation efficiency were compared between the two groups.  Results  After 4 batches of training, the Kappa value between the model's decisions and labels improved from 0.610 to 1.000. The consistency with nurse decisions increased from 0.487 to 0.726, which was higher than the inter-nurse consistency (Kappa=0.612). Using the agreement of two nurses as the standard, the model's accuracy for real cases improved from 56.00% to 80.95%, and the weighted F1 score increased from 0.40 to 0.72. Using the agreement of one nurse as the standard, the accuracy and precision reached 100.00% after training. In clinical validation, 96.55% (56/58) of the model's decisions were recognized by nurses. Compared to the control group, the model group showed no significant differences in blood pressure control during dialysis (χ²=5.744, P=0.057), blood pressure control after dialysis (χ²=0.747, P=0.688), headaches during dialysis (P＞0.999), cramps (χ²=0.347, P=0.541), and manifestations of disequilibrium syndrome (P=0.575). However, the assessment time before machine operation was significantly shortened (t=45.300, P＜0.001).  Conclusion   The large language model optimized with real cases can achieve accurate and rapid decision-making for MHD dialysis modes, demonstrating good clinical applicability and potential for wider adoption.

Key words: End-stage renal disease, Maintenance hemodialysis, Auxiliary decision-making model, Intelligent dialysis mode, Large language model