AI与药学实践 第8期(总第38期):Journal of the American Medical Informatics Association

来源：Journal of the American Medical Informatics Association (JAMIA)

AI与药学实践 第8期（总第38期） 

——————————————————————————————

本期聚焦《JAMIA》期刊中医学信息学与药学实践交叉领域的最新研究。精选5篇高质量论文，涵盖临床文本用药信息NLP提取、ML预测心衰患者药物依从性、ICU数字孪生与用药管理、电子药物相互作用预警有效性系统综述，以及大语言模型在儿科抗生素管理处方识别中的应用等前沿方向。

▎论文目录·Table of Contents

——————————————————————————————

▎综述·Reviews

Effect of electronic drug-drug interaction alerts on patient and clinician outcomes: a systematic review

Journal of the American Medical Informatics Association (JAMIA) 丨 2025-10-01 丨 Systematic Review

▍电子药物相互作用预警对患者和临床医生结局的影响：系统综述

研究背景

药物相互作用临床决策支持系统（CDSS-DI）已在医院信息系统广泛部署，但其临床有效性长期缺乏系统性评估。该综述全面评估了CDSS-DI对患者和临床医生结局的真实影响，对用药安全信息化建设具有重要指导意义。

核心发现

• 纳入8项研究共43413名患者，涵盖RCT和前瞻性队列设计

• CDSS-DI预警未显著降低死亡率（OR: 1.94）或不良药物相互作用发生率（OR: 0.86）

• CDSS-DI对处方行为仅有轻微影响（OR: 2.08, P = 0.05）

• 系统呼吁优化预警准确性、临床相关性和患者特异性整合

领域意义

该系统综述首次全面评估了药物相互作用预警系统的临床有效性，揭示了当前CDSS-DI对患者结局改善有限的问题，对推动用药安全决策支持系统的智能化升级和精准预警具有重要指导价值。

英文摘要 (Abstract)

BACKGROUND: Drug interaction checking software within clinical decision support systems (CDSS-DI) is widely used, but its patient relevance and accuracy are variable, and its impact on patient outcomes has not been demonstrated. This review aimed to compare the effectiveness of using CDSS-DI versus usual care without CDSS-DI. MATERIALS AND METHODS: We searched multiple databases from 1990 to the end of 2024, including randomized controlled trials (RCTs) or prospective cohort studies evaluating CDSS-DI at the prescribing, dispensing, or administration stages with a control group and assessing clinical, process, or burden outcomes. Risk of bias was assessed using Cochrane RoB 2 and ROBINS-I. Random-effects meta-analysis was performed using the meta 6.0 package in R. RESULTS: Eight studies were included: 7 RCTs (2 parallel, 5 cluster) and 1 prospective cohort study, with a total of 43 413 patients. Mortality was not significantly different between intervention (0.14%) and control groups (0.07%) (OR: 1.94 [95% CI: 0.88-4.29], P = 0.078). One study reported a 3-hour reduction in length of stay (P = 0.0021), though clinical significance may be limited. CDSS-DI alerts had a modest effect on prescribing behavior (OR: 2.08 [95% CI: 1.01-4.27], P = 0.05) but did not significantly reduce target adverse drug interactions (OR: 0.86 [95% CI: 0.56-1.34], P = 0.37). DISCUSSION: Surprisingly few high-quality studies have evaluated the impact of CDSS-DI on patient or clinician outcomes. The current evidence suggests CDSS-DI has no significant benefit on patient-important outcomes. Given potential harms and important time burden, CDSS-DI alerts need improvement. CONCLUSION: CDSS-DI alerts have not shown significant improvement in patient-important outcomes. Optimizing alert accuracy, clinical relevance, and patient-specific integration is critical to enhancing their value in clinical practice.

DOI: 10.1093/jamia/ocaf139

——————————————————————————————

▎研究论文·Research Articles

Efficient extraction of medication information from clinical notes: an evaluation in 2 languages

Journal of the American Medical Informatics Association (JAMIA) 丨 2025-12-01 丨 Research Article

▍从临床笔记中高效提取用药信息：两种语言的评估研究

研究背景

临床笔记中包含大量与患者用药方案相关的非结构化文本信息，传统人工提取效率低下。基于Transformer的NLP方法在英文临床文本中已有应用，但跨语言移植和计算成本仍是实际部署中的关键挑战。

核心发现

• 提出了一种新型Transformer架构，可从临床文本中端到端提取用药信息

• 在法语和英语数据集上F值分别达到0.82和0.96，与最先进方法相当

• 计算成本降低10倍，适合资源有限的医院IT环境

• 验证了方法在跨语言场景（法语和英语）中的可移植性

领域意义

该研究为医院药学信息系统提供了高效的用药信息自动提取方案，显著降低了部署门槛和计算成本，对推动NLP技术在多语言医院环境中的实际应用具有重要意义。

英文摘要

OBJECTIVE: Evaluate the accuracy, computational cost, and portability of a novel natural language processing (NLP) approach for extracting medication information from clinical narrative texts. MATERIALS AND METHODS: We propose a transformer-based original architecture for extracting entities and their relations related to patient medication regimens. First, we train and evaluate the model on French clinical notes using a newly annotated French corpus from Strasbourg University Hospital. Second, we assess portability by evaluating on English clinical documents from the 2018 n2c2 shared task. We evaluate extraction accuracy and computational cost by comparing to existing transformer-based approaches. RESULTS: The proposed architecture achieves performance comparable to state-of-the-art methods on the relation extraction task, with F-scores of 0.82 and 0.96 on French and English datasets respectively (compared to 0.81 and 0.95 for existing methods), while reducing computational cost by a factor of 10. End-to-end (NER + relation extraction) F1 performance was 0.69 and 0.82 on French and English corpora respectively. DISCUSSION: While existing systems developed for English notes can be deployed in French hospital settings with reasonable effort, we find that an alternative architecture can achieve comparable extraction performance, lower computational cost, and support end-to-end medication information extraction from both French and English clinical texts. CONCLUSION: The proposed architecture can extract medication information from clinical text with high performance and low computational cost, making it suitable for hospital IT environments which are typically resource-limited.

DOI: 10.1093/jamia/ocaf113

——————————————————————————————

Machine learning based prediction of medication adherence in heart failure using large electronic health record cohort with linkages to pharmacy-fill and neighborhood-level data

Journal of the American Medical Informatics Association (JAMIA) 丨 2025-12-01 丨 Research Article

▍基于机器学习的心力衰竭患者药物依从性预测：利用链接药房配药和社区层面数据的大型电子健康记录队列研究

研究背景

心力衰竭患者药物依从性差会导致预后不良，及时识别不依从高风险患者并实施干预至关重要。现有方法多依赖单一数据源，缺乏整合电子健康记录、药房配药数据和社区层面多维度信息的综合预测模型。

核心发现

• 基于34697名心力衰竭患者的大型EHR-药房配药数据队列，纳入超过120个多维度预测因子

• 超级学习者（superlearner）集成模型MAE最低（18.9%），显著优于传统OLS回归

• 首次将WHO不依从障碍模型的多层面因素（患者、治疗、医疗、社区）整合入ML预测框架

• 模型可识别可从药物依从性干预中获益的HF患者，为临床精准干预提供依据

领域意义

该研究建立了整合EHR、药房配药和社区数据的机器学习预测框架，为临床药师识别药物不依从高风险患者提供了实用工具，对推动精准药学服务和个体化用药管理具有重要价值。

英文摘要

OBJECTIVE: Timely interventions can improve medication adherence, yet identifying patients at risk for nonadherence remains challenging in clinical settings. This study aimed to develop and validate flexible machine learning (ML) models for predicting adherence to guideline-directed medical therapy (GDMT) in heart failure (HF). MATERIALS AND METHODS: We used a large electronic health record (EHR) cohort of 34 697 HF patients seen at NYU Langone Health between April 1, 2021 and October 31, 2022, all with active prescriptions for >=1 GDMT. The outcome was 6-month GDMT adherence measured by proportion of days covered (PDC). Using the WHO nonadherence barrier model, over 120 predictors spanning patient-level, therapy-level, health care-level, and community-level factors were included. We compared multiple ML models and their ensemble (superlearner) to traditional regression (OLS) on mean absolute error (MAE) from 10-fold cross-validation, percent improvement in MAE relative to superlearner, and predicted differences across PDC deciles. RESULTS: The flexible nonparametric predictor superlearner performed best. Superlearner and quantile random forest had the lowest MAE (mean [95% CI] 18.9% [18.7%-19.1%] for both), followed by quantile neural networks (19.5% [19.3%-19.7%]) and kernel support vector regression (19.8% [19.6%-20.0%]). Gradient boosting and OLS performed worst, with MAEs 17% and 14% higher than superlearner respectively. Superlearner showed superior predicted differences across deciles. CONCLUSION: This development-phase study suggests that integrating EHR-pharmacy data with machine learning can identify HF patients who may benefit from medication adherence interventions.

DOI: 10.1093/jamia/ocaf162

——————————————————————————————

Toward digital twins in the intensive care unit: a medication management case study

Journal of the American Medical Informatics Association (JAMIA) 丨 2025-10-01 丨 Research Article

▍迈向重症监护病房的数字孪生：一项用药管理案例研究

研究背景

数字孪生技术在医疗领域的应用尚处于早期阶段。ICU用药管理具有高度复杂性和专科特异性，将大语言模型与数字孪生概念结合，有望为重症患者提供个性化用药决策支持。

核心发现

• 首次将LLaMA-3结合LoRA微调应用于ICU用药管理数字孪生体开发

• 内科ICU专科特异性训练模型BERTScore达0.842，显著优于零样本基线

• 验证了数字孪生技术在重症用药管理场景中的可行性

• 揭示了大语言模型专科特异性微调对临床决策支持的重要价值

领域意义

该研究开创性地将数字孪生概念与大型语言模型结合用于ICU用药管理，为精准药学在重症医学中的应用提供了新思路，标志着AI驱动的个性化临床药学决策支持进入新阶段。

英文摘要

OBJECTIVE: Evaluate the effectiveness of a digital twin, developed using LLaMA-3 and fine-tuned with LoRA on ICU physician notes from the Medical Information Mart for Intensive Care III (MIMIC-III) dataset, where medication information is masked to create training and testing datasets. We assess whether domain-specific training improves treatment recommendation accuracy compared to other ICU specialties or zero-shot baselines. MATERIALS AND METHODS: The digital twin is created using LLaMA-3, fine-tuned on discharge summaries from MIMIC-III where medication information was masked for training and testing. Evaluation uses 1000 internal medicine ICU notes, with performance measured by BERTScore and ROUGE-L. A zero-shot baseline relying solely on in-context instructions without training was also evaluated. RESULTS: The model fine-tuned on internal medicine ICU notes achieved the highest BERTScore (0.842), outperforming models trained on other specialties or mixed datasets. The zero-shot model performed worst, highlighting the necessity of domain-specific training. DISCUSSION: Results show that domain-specific training significantly improves the digital twin's accuracy for treatment recommendations compared to general or zero-shot approaches. Tailoring models to specific ICU domains can enhance their utility in clinical decision support. CONCLUSION: Context-specific fine-tuning of LLMs is essential for developing effective digital twins, providing foundational insights for personalized clinical decision support.

DOI: 10.1093/jamia/ocaf127

——————————————————————————————

A SNAPpy use of large language models: using large language models to classify treatment plans in pediatric acute otitis media

Journal of the American Medical Informatics Association (JAMIA) 丨 2025-12-01 丨 Research Article

▍大语言模型在儿科急性中耳炎治疗方案分类中的应用：SNAP处方识别研究

研究背景

急性中耳炎是儿童抗生素过度使用的主要原因，安全网抗生素处方（SNAPs）作为推荐的抗生素管理策略，因缺乏结构化文档记录而难以大规模追踪和评估。大语言模型为从非结构化临床笔记中自动识别处方模式提供了新途径。

核心发现

• GPT-4o零样本提示在AOM治疗方案分类中准确率达97.8%

• LLM Versa性能优于本地微调的Clinical-Longformer模型（93.3%）

• 基于5707次就诊记录的大规模验证，374份经人工审核

• 为儿科抗生素管理中的处方模式追踪提供了成本高效的自动化工具

领域意义

该研究展示了大语言模型在处方实践模式识别和抗生素管理中的实用价值，为临床药师和药物管理人员利用AI技术开展处方审核和抗菌药物管理提供了新的技术路径。

英文摘要

BACKGROUND AND SIGNIFICANCE: Acute otitis media (AOM) is the leading driver of antibiotic overuse in children. Safety-net antibiotic prescriptions (SNAPs) are a recommended antibiotic stewardship strategy, but they are difficult to identify due to lack of structured documentation. OBJECTIVE: This study validates the accuracy of a HIPAA-compliant GPT-4o-based large language model (LLM), Versa, for classifying AOM treatment plans from physician notes. METHODS: A retrospective cross-sectional study analyzed pediatric AOM encounters. Treatment plans were classified using various prompting strategies, with a representative sample manually reviewed by 2 pediatricians as a gold standard. A locally fine-tuned Clinical-Longformer model was also trained and tested, with performance compared to Versa and manual review. RESULTS: The study included 5707 encounters, of which 374 were manually reviewed. Zero-shot prompting achieved 97.8% accuracy; few-shot prompting achieved 85% accuracy. The Clinical-Longformer achieved 93.3% accuracy. CONCLUSION: Versa can effectively identify AOM treatment plans, providing a cost-efficient tool for tracking prescribing practice patterns in pediatric antibiotic stewardship efforts.

DOI: 10.1093/jamia/ocaf170

——————————————————————————————

▎参考文献·References

[1] Thibaut Fabacher, Erik-André Sauleau, Emmanuelle Arcay, Bineta Faye, Maxime Alter, Archia Chahard, Nathan Miraillet, Adrien Coulet, Aurélie Névéol. Efficient extraction of medication information from clinical notes: an evaluation in 2 languages. J Am Med Inform Assoc. 2025-12-01. https://pubmed.ncbi.nlm.nih.gov/41071911/

[2] Samrachana Adhikari, Tyrel Stokes, Xiyue Li, Yunan Zhao, Cassidy Fitchett, Nathalia Ladino, Steven Lawrence, Min Qian, Young S Cho, Carine Hamo, John A Dodson, Rumi Chunara, Ian M Kronish, Amrita Mukhopadhyay, Saul B Blecker. Machine learning based prediction of medication adherence in heart failure using large electronic health record cohort with linkages to pharmacy-fill and neighborhood-level data. J Am Med Inform Assoc. 2025-12-01. https://pubmed.ncbi.nlm.nih.gov/41032036/

[3] Behnaz Eslami, Majid Afshar, Samie Tootooni, Timothy A Miller, Matthew M Churpek, Yanjun Gao, Dmitriy Dligach. Toward digital twins in the intensive care unit: a medication management case study. J Am Med Inform Assoc. 2025-10-01. https://pubmed.ncbi.nlm.nih.gov/40981526/

[4] Anne M Holbrook, Jessyca Matos Silva, Junaid Ahmed Yaser Faruque, Jiawen Deng, Tyler Schneider, Aliya Jaffer. Effect of electronic drug-drug interaction alerts on patient and clinician outcomes: a systematic review. J Am Med Inform Assoc. 2025-10-01. https://pubmed.ncbi.nlm.nih.gov/40853269/

[5] Jessica J Pourian, Ben Michaels, Anh Vo, A Jay Holmgren, Augusto Garcia-Agundez, Valerie Flaherman. A SNAPpy use of large language models: using large language models to classify treatment plans in pediatric acute otitis media. J Am Med Inform Assoc. 2025-12-01. https://pubmed.ncbi.nlm.nih.gov/41069025/

——————————————————————————————

关注我们，追踪全球药学前沿

第8期 · 总第38期 丨 2026年7月