P-1305. Identifying Animal Exposures from Clinical Documents Using Large Language Models

P-1305. 利用大型语言模型从临床文档中识别动物暴露史

阅读:1

Abstract

BACKGROUND: Information about animal exposure is crucial to understanding zoonotic infectious diseases; however, it is typically unavailable in structured clinical data. We evaluated the automated extraction of animal exposure mentions from clinical notes and summarized initial observations. [Figure: see text] METHODS: Clinical notes from the Department of Veterans Affairs (VA) data were extracted if associated with indicators for diseases which included CDC Nationally Notifiable Diseases (NND) as well as early cases of COVID-19 and mpox (i.e., positive labs, chart review, or diagnosis codes). These were then processed into text snippets containing animal keywords (e.g., cats, cattle, birds, etc.) and given to annotators who then assigned a category of: affirmed, denied/negated, or no exposure, including pets and farm animals. To augment the size of our training set, large language models (LLM) were also used to generate snippets for each category. A machine learning classification model was trained using a LLM and few-shot learning. After validation, this model was used to infer animal exposure mentions among infectious disease cases which were not part of annotation to examine the distribution of animals mentioned. [Figure: see text] RESULTS: The model’s accuracy on the held out test set of 115 (40%) text instances was 90% PPV and 91% sensitivity. Example snippets and annotations are shown in Figure 1. Following validation, the model was applied to 82,937 documents related to a selection of disease categories. The distribution of classified exposures among these documents was 2.7% affirmed, 0.1% denied/negated, and 97.2% no exposure. Figure 2 shows that prevalence of affirmed exposure varied by disease where non-zoonotic diseases such as COVID-19 were relatively lower. Some top animals terms in affirmed exposures are shown in Figure 3. Figure 4 is a distribution of top terms affirmed in tularemia cases. [Figure: see text] CONCLUSION: Automated extraction of animal exposure is feasible with small, manually generated training sets and acceptable accuracy. Our approach summarized exposures documented among several categories of infectious disease. While this work was performed retrospectively, it has been used to enhance case-finding for the biosurveillance of zoonotic diseases. [Figure: see text] DISCLOSURES: All Authors: No reported disclosures

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。