放射学中人工智能工具的成本效益:一项系统综述
本系统综述检索截至2025年1月23日的五个数据库,纳入10项对放射学AI诊断成像干预进行正式经济学分析的研究,其中9项采用质量调整生命年(QALY)的成本-效用分析,1项采用伤残调整生命年(DALY)。所有研究均为理论建模(马尔可夫、决策树或混合模拟),无前瞻性真实世界成本效益数据,多数基于美国或英国的公开医疗数据,结论为在模型特定假设和支付意愿阈值下AI可能具有成本效益。方法学异质性妨碍了Meta分析。
要点
- 纳入10项研究均为模型-based经济学评价,无前瞻性真实世界成本效益数据。
- 9项研究采用QALY成本-效用分析,1项采用DALY分析。
- 多数研究基于美国或英国公开医疗数据,结论为在模型特定假设和支付意愿阈值下AI可能具有成本效益。
核心结论
- 当前放射学AI成本效益证据有限、完全基于模型,且缺乏真实世界前瞻性数据验证;未来研究应采用标准化评价框架并纳入实证临床数据。
研究设计与样本
- 系统综述,PRISMA合规检索PubMed、Cochrane Library、Scopus、Web of Science和Embase截至2025年1月23日的英文文献,纳入对放射学AI诊断成像干预进行正式经济学分析的研究,使用CHEERS-AI清单评价。
诊断标准与关键数值
- 共识别360篇文献,10篇符合纳入标准。
- 9项研究采用质量调整生命年(QALY)的成本-效用分析,1项采用伤残调整生命年(DALY)分析。
- 所有研究均采用理论建模(马尔可夫、决策树或混合模拟),无前瞻性真实世界成本效益数据。
局限性
- 方法学异质性妨碍了Meta分析。
- 当前证据完全基于模型,缺乏前瞻性真实世界成本效益数据验证。
- 多数研究使用美国或英国的公开医疗数据建模,结论基于模型特定假设和支付意愿阈值。
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
OBJECTIVES: To systematically review the evidence on the cost-effectiveness of artificial intelligence (AI) interventions for diagnostic imaging in radiology, thereby identifying methodological gaps and priorities for future research.
MATERIALS AND METHODS: A PRISMA-compliant search of PubMed, Cochrane Library, Scopus, Web of Science, and Embase was conducted for English-language studies published until January 23, 2025. Eligible studies evaluated AI-based radiology interventions for diagnostic imaging with formal economic analysis. Studies were appraised using the Consolidated Health Economic Evaluation Reporting Standards for Artificial Intelligence-based Interventions (CHEERS-AI) checklist.
RESULTS: Of 360 publications identified, ten met the inclusion criteria. Nine studies reported cost-utility analyses using quality-adjusted life-years (QALYs) and one used disability-adjusted life-years (DALYs). All studies employed theoretical modeling (Markov, decision-tree, or hybrid simulations), with no prospective real-world cost-effectiveness data. Applications included cancer screening, acute stroke detection, infection control, and opportunistic detection of incidental findings. Most included studies used publicly available healthcare data from the United States or the United Kingdom to model cost-effectiveness outcomes, and concluded that AI may be cost-effective under model-specific assumptions and willingness-to-pay thresholds. Methodological heterogeneity precluded meta-analysis.
CONCLUSION: Current evidence on the cost-effectiveness of AI in radiology is limited, model-based, and lacks validation with real-world prospective data. Future research should employ standardized evaluation frameworks and incorporate empirical clinical data to better inform implementation decisions.
KEY POINTS: Question Economic evaluations are important for sustained implementation of AI in radiology, but current evidence on cost-effectiveness is limited. Findings Ten studies, predominantly from the United States and the United Kingdom, used model-based analyses and generally indicated cost-effectiveness under study-specific assumptions, without prospective real-world cost data. Clinical relevance Despite increasing adoption of AI tools in radiology, current cost-effectiveness evidence is limited, wholly model-based, and methodologically heterogeneous. Standardized economic evaluations using prospective real-world data are needed to support informed adoption decisions.
引用
- Brin D, Tau N. Cost-effectiveness of artificial intelligence tools in radiology: a systematic review. Eur Radiol. 2025. PMID: 41441999. doi:10.1007/s00330-025-12242-4
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。