The effects of multitype prompt engineering for large language models in hypertension treatment decisions

作者：Zeyan Li, Henyang Liu, Wuping Tan, Du Tang, Shoupeng Duan, Bowen Zhou, Long Tang, Xuyang Hu, Liying Huang, Peng Zhao, Wenqiang Fang, Bing Wu, Jinjun Liu, Yijun Wang, Jun Wang · 发表于：npj Digital Medicine · 年份：2026 · DOI：10.1038/s41746-026-02645-y · 被引用次数：2 · 研究领域：Machine Learning in Healthcare、Artificial Intelligence in Healthcare and Education、Topic Modeling

The effects of various prompt engineering on Large Language Models (LLMs) performance in hypertension decision-making are not yet fully understood. We evaluate the impact of different prompt engineering on LLM performance in hypertension treatment decision-making. We conducted a two-stage validation study using 300 de-identified simulated hypertension cases based on real-world clinical scenarios. ChatGPT-4.1 with Guidance-Self-Consistency achieved optimal performance (91.3% accuracy), nearing expert-level competency, while zero-shot prompting yielded worst results (62.7% with DeepSeek-V3). Optimal LLM assistance consistently enhanced physicians' average accuracy across all levels (community hospital: 73.4% to 82.5%; county hospital: 84.0% to 87.9%; teaching hospital: 91.5% to 92.0%) and reduced inappropriate regimen rates. The worst LLM configurations decreased physician performance below baseline, increasing inappropriate regimen rates from 26.6% to 35.2% across all levels. Effectively designed prompt strategies enable LLMs to provide reliable hypertension treatment recommendations, thereby supporting physicians' clinical decisions. This study has been trial-registered (ChiCTR2500099307, March 21, 2025).