Scholarlink

This study proposes a robust framework for automated discharge medication recommendation by leveraging LoRA-based fine-tuning of large language models (LLMs) on Chinese electronic health records (EHRs), addressing the CHIP 2025 shared task. We introduce a multi-stage framework that integrates several key techniques: (1) parameter-efficient instruction fine-tuning of four base LLMs (Qwen3-8B, Qwen3-4B-Instruct-2507, Llama-3.1-8B-Instruct, and Mistral-7B-Instruct-v0.3) using Low-Rank Adaptation (LoRA); (2) stabilization of predictions via repeated sampling and ensemble voting to mitigate generative variability and model-specific biases; and (3) systematic output refinement through pseudo-labeling and rule-based post-processing for drug name normalization. Experiments on 3,602 training and 1,722 test samples demonstrate substantial improvements over baseline performance. The final ensemble model achieved a Jaccard score of 0.4870, an F1 score of 0.6014, and a final evaluation score of 0.5442, validating the effectiveness of the proposed framework in enhancing the accuracy and practical reliability of medication recommendation systems.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LoRA-Fine-Tuned LLMs for Discharge Medication Recommendation on Chinese EHRs

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xiao Su

摘要

This study proposes a robust framework for automated discharge medication recommendation by leveraging LoRA-based fine-tuning of large language models (LLMs) on Chinese electronic health records (EHRs), addressing the CHIP 2025 shared task. We introduce a multi-stage framework that integrates several key techniques: (1) parameter-efficient instruction fine-tuning of four base LLMs (Qwen3-8B, Qwen3-4B-Instruct-2507, Llama-3.1-8B-Instruct, and Mistral-7B-Instruct-v0.3) using Low-Rank Adaptation (LoRA); (2) stabilization of predictions via repeated sampling and ensemble voting to mitigate generative variability and model-specific biases; and (3) systematic output refinement through pseudo-labeling and rule-based post-processing for drug name normalization. Experiments on 3,602 training and 1,722 test samples demonstrate substantial improvements over baseline performance. The final ensemble model achieved a Jaccard score of 0.4870, an F1 score of 0.6014, and a final evaluation score of 0.5442, validating the effectiveness of the proposed framework in enhancing the accuracy and practical reliability of medication recommendation systems.