自然语言处理在多队列中客观识别肝性脑病
本研究利用自然语言处理(NLP)从肝硬化住院病历中提取与肝性脑病(HE)相关的术语,在推导队列及两个验证队列中评估其识别HE的性能。外部验证队列145例(6547份记录,35% HE诊断)中敏感性为94.1%、NPV为84.2%。机器学习交叉验证显示两个验证队列中排名前5的特征与推导队列相似,且术语组合优于单个术语。
要点
- NLP可从住院病历中提取5个简单术语(扑翼样震颤、精神状态改变、意识模糊、乳果糖/利福昔明使用)定义HE。
- 5个术语组合在推导队列敏感性85.8%、NPV 82.1%,内部验证队列敏感性和NPV均为100%,外部验证队列敏感性94.1%、NPV 84.2%。
- 术语组合对HE定义的贡献优于单个术语,机器学习交叉验证显示验证队列前5特征与推导队列相似。
核心结论
- 通过NLP从住院病历中提取的5个简单短语(扑翼样震颤、精神状态改变、意识模糊、开始/继续乳果糖/利福昔明)可在三个队列中客观定义HE。
研究设计与样本
- 研究采用NLP术语选择(基于指南回顾和医师评估),在推导队列(Richmond VA医疗中心肝硬化住院患者,MELD 3.0≥15)中筛选高敏感性/高NPV术语,并在两个独立验证队列(Richmond VA和Virginia Commonwealth University)中验证。
- 采用机器学习交叉验证研究各术语对HE诊断的独立贡献。
主要发现
- 推导队列中5个变量与HE相关,敏感性85.8%、NPV 82.1%。
- 内部验证队列中5个术语的敏感性和NPV均为100%。
- 外部验证队列中敏感性94.1%、NPV 84.2%。
- 交叉验证机器学习显示两个验证队列中排名前5的特征与推导队列相似。
- 术语组合对HE定义的贡献优于单个术语。
临床怎么用
- 该NLP方法可客观识别住院HE事件,可能有助于未来肝移植优先政策制定和临床试验中HE事件裁定。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
INTRODUCTION: Despite being a major clinical, financial, and psychosocial burden, hepatic encephalopathy (HE) is not specifically prioritized for liver transplant. Moreover, subjectivity in HE definition can hinder outcome analysis for multicenter trials. We used natural language processing (NLP) to determine terms associated with an inpatient HE episode across derivation and 2 validation cohorts. METHODS: NLP terms related to HE were selected using guideline review and physician evaluation. Derivation cohort: Charts from cirrhosis inpatients at Richmond Veterans Affairs (VA) Medical Center with MELD 3.0 ≥ 15 underwent NLP. Terms with highest sensitivity/negative predictive value (NPV) were then validated in 2 independent cohorts from North American Consortium of Study of End-stage Liver Disease (cohorts in Richmond VA and Virginia Commonwealth University). Individual contribution of terms toward HE diagnosis was studied with machine-learning cross-validation. RESULTS: Derivation cohort consisted of four hundred thirty-two Veterans (4,048 notes, 34% HE diagnosis) showed 5 variables associated with HE (asterixis, altered mental status, confusion, initiation/continuation of lactulose/rifaximin) with 85.8%-sensitivity/82.1%-NPV. One hundred sixty-four patients (1,536 notes, 26% HE diagnosis) were included for internal validation, where the sensitivity/NPV of these 5 terms was 100% each. Overall, 145 patients (6,547 notes, 35% HE) were included with 94.1% sensitivity and 84.2%-NPV in an external validation cohort. On cross-validated machine learning, the top 5 features in both validation cohorts were similar to the derivation cohort. Contribution of the combination of these terms was superior to individual terms in defining HE. DISCUSSION: Five signs, symptoms, and treatment decisions with simple phrases extracted from inpatient charts using NLP define HE across 3 cohorts. This objectivity could aid future policy on liver transplant priority and HE-event adjudication in clinical trials.
引用
- Silvey S, Patel N, Davis BC, et al. Use of Natural Language Processing to Objectively Identify Hepatic Encephalopathy in Multiple Cohorts. Am J Gastroenterol. 2025. PMID: 41196122. doi:10.14309/ajg.0000000000003824
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。