人工智能辅助适应证选择对影像临床决策支持中适宜性评分的影响
该多中心研究分析了AI工具用于自由文本适应证分析前后影像订单的适宜性评分与结构化录入情况,发现部署后评分订单从30%升至52%,结构化适应证订单从34.6%升至67.3%,但仍有47.8%订单未评分,其中62.7%为纯自由文本适应证。
要点
- AI辅助适应证分析后,影像订单被评分的比例从30%升至52%,结构化适应证订单从34.6%升至67.3%。
- 部署后仍有47.8%订单未评分,其中62.7%为纯自由文本适应证。
- 非医师开单者订单被评分的可能性低于医师(OR 0.80)。
核心结论
- 在影像临床决策支持中嵌入AI辅助与结构化适应证订单增加相关,并独立预测更高评分可能性,但48%订单仍未评分,原因包括开单者行为和基础设施障碍。
研究设计与样本
- 多中心医疗系统研究,分析AI工具部署前7个月(2020年3月1日至2020年9月21日)115,079份订单和部署后(2020年10月20日至2021年5月13日)150,950份订单,使用χ2检验和多变量logistic回归(bootstrap法校正协变量)。
诊断标准与关键数值
- 临床决策支持评分分为不适当、可能适当、适当或未评分四类。
- 适应证类型分为结构化、自由文本、两者兼有或无。
- 多变量分析中,AI工具部署后订单被评分的OR为2.7(95%CI 2.63-2.78,P<.001)。
- 非医师开单者订单被评分的OR为0.80(95%CI 0.78-0.83,P<.001),低于医师。
- MR订单被评分的OR为0.84(95%CI 0.82-0.87),PET订单为0.12(95%CI 0.10-0.13),均低于CT(P<.001)。
临床应用与阅片要点
- AI辅助自由文本适应证分析可提高影像订单的结构化录入率和适宜性评分覆盖率。
- 纯自由文本适应证和特定检查类型(如PET、MR)及非医师开单者订单更易未被评分,阅片时需留意评分缺失情况。
局限性
- 原文未报告局限性。
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)。
英文原文(摘要)· 点击展开
PURPOSE: The aim of this study was to assess appropriateness scoring and structured order entry after the implementation of an artificial intelligence (AI) tool for analysis of free-text indications.
METHODS: Advanced outpatient imaging orders in a multicenter health care system were recorded 7 months before (March 1, 2020, to September 21, 2020) and after (October 20, 2020, to May 13, 2021) the implementation of an AI tool targeting free-text indications. Clinical decision support score (not appropriate, may be appropriate, appropriate, or unscored) and indication type (structured, free-text, both, or none) were assessed. The χ2 and multivariate logistic regression adjusting for covariables with bootstrapping were used.
RESULTS: In total, 115,079 orders before and 150,950 orders after AI tool deployment were analyzed. The mean patient age was 59.3 ± 15.5 years, and 146,035 (54.9%) were women; 49.9% of orders were for CT, 38.8% for MR, 5.9% for nuclear medicine, and 5.4% for PET. After deployment, scored orders increased to 52% from 30% (P < .001). Orders with structured indications increased to 67.3% from 34.6% (P < .001). On multivariate analysis, orders were more likely to be scored after tool deployment (odds ratio [OR], 2.7, 95% CI, 2.63-2.78; P < .001). Compared with physicians, orders placed by nonphysician providers were less likely to be scored (OR, 0.80; 95% CI, 0.78-0.83; P < .001). MR (OR, 0.84; 95% CI, 0.82-0.87) and PET (OR, 0.12; 95% CI, 0.10-0.13) were less likely to be scored than CT (; P < .001). After AI tool deployment, 72,083 orders (47.8%) remained unscored, 45,186 (62.7%) with free-text-only indications.
CONCLUSIONS: Embedding AI assistance within imaging clinical decision support was associated with increased structured indication orders and independently predicted a higher likelihood of scored orders. However, 48% of orders remained unscored, driven by both provider behavior and infrastructure-related barriers.
引用
- Shreve, Fried, Liu, et al. Impact of Artificial Intelligence-Assisted Indication Selection on Appropriateness Order Scoring for Imaging Clinical Decision Support. J Am Coll Radiol. 2023. PMID: 37390881. doi:10.1016/j.jacr.2023.04.016
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。