The progress in natural language processing (NLP) using large language models (LLMs) has greatly improved patient information extraction from clinical narratives. However, most methods based on the fine-tuning strategy have limited transfer learning ability for cross-domain applications. This study proposed a novel approach that employs a soft prompt-based learning architecture, which introduces trainable prompts to guide LLMs toward desired outputs. We examined two types of LLM architectures, including encoder-only GatorTron and decoder-only GatorTronGPT, and evaluated their performance for the extraction of social determinants of health (SDoH) using a cross-institution dataset from the 2022 n2c2 challenge and a cross-disease dataset from the University of Florida (UF) Health. The results show that decoder-only LLMs with prompt tuning achieved better performance in cross-domain applications. GatorTronGPT achieved the best F1 scores for both datasets, outperforming traditional fine-tuned GatorTron by 8.9% and 21.8% in a cross-institution setting, and 5.5% and 14.5% in a cross-disease setting.

翻译：自然语言处理（NLP）在大型语言模型（LLMs）应用方面的进展极大提升了从临床叙述中提取患者信息的能力。然而，基于微调策略的大多数方法在跨领域应用中泛化能力有限。本研究提出了一种新方法，采用基于软提示的学习架构，通过引入可训练提示来引导LLMs生成期望输出。我们研究了两种类型的LLM架构，包括仅编码器GatorTron和仅解码器GatorTronGPT，并使用2022年n2c2挑战赛的跨机构数据集和佛罗里达大学（UF）健康中心的跨疾病数据集，评估了它们在提取健康社会决定因素（SDoH）方面的性能。结果表明，采用提示调优的仅解码器LLM在跨领域应用中表现更佳。GatorTronGPT在两个数据集上均取得了最佳F1分数，在跨机构设置中分别比传统微调GatorTron高出8.9%和21.8%，在跨疾病设置中分别高出5.5%和14.5%。