跳到正文
原文
Nature Medicine·· 2026-10-14AI 评分45

Nature Medicine 特刊:医疗人工智能部署的挑战与经验

Making AI work in healthcare

AI 导读

Nature Medicine 本期聚焦医疗人工智能部署挑战,指出当前 AI 工具在临床中的实际影响证据有限。研究显示,美国超过 90% 的医疗系统已部署 AI 工具,但仅 53% 或更低比例认为其高度有用。文章强调需通过前瞻性研究验证 AI 对临床结局的影响,并建议建立部署前后的评估框架,如美国资助的 6270 万美元 ADVOCATE 项目。

正文

Drawing on experiences across diverse settings, this issue of Nature Medicine explores the challenges of and lessons learned from the deployment of artificial intelligence models into healthcare.

Advances in machine learning and generative artificial intelligence (AI) have produced tools that can potentially support an ever-expanding range of clinical tasks, from the interpretation of medical images to assisting clinicians in documenting and making decisions about patient care. Thus, AI technologies are being deployed into healthcare at a pace rarely seen for new medical innovations. In the USA, the Food and Drug Administration has begun a pilot program1 for deploying AI clinical decision support tools before they are authorized, while over 800 healthcare institutions2,3 have already deployed China’s DeepSeek model after its release. Yet a gap remains between AI hype and clinical reality.

Evidence remains limited on whether the deployment of these tools is having any real impact on patients, healthcare providers or health systems. A survey has already found that over 90% of US health systems have begun deploying AI tools for tasks such as radiology and clinical documentation, and 53% or less found these tools to be highly useful4.

Many of these implementation challenges stem from a disconnect between how AI is evaluated in research and how it is ultimately used in clinical practice. AI models often achieve impressive performance metrics in retrospective evaluation and reader studies, but these results do not necessarily translate into clinical value unless they are integrated into clinical workflows. Drawing on their experience developing an AI eye-care system in China, Yan et al. found that the AI technology was feasible only when embedded at every stage of the care pathway, emphasizing that successful AI deployment depends on more than model performance alone.

A related challenge is that different healthcare systems vary considerably in their resources, readiness of data infrastructure, models of care and workforce capacity. As a result, technologies that are feasible in one hospital setting may need to be substantially adapted to be implemented effectively elsewhere. Kabukye et al. highlight this challenge in low- and middle-income countries, where limited computational resources and shortages of specialists mean that AI systems will need to be redesigned as ‘frugal’ solutions tailored to local needs. Adaptation of medical AI to different healthcare contexts is especially important when considering international scale up. For example, Tiwari et al. report that deploying the same diabetic retinopathy AI system across India, Thailand and Australia required vastly different implementation strategies shaped by local healthcare priorities and clinical workflows.

AI systems already develop limitations in evaluation studies. For example, large language models have been shown to generate inaccurate and misleading medical advice in research contexts, and real-world safety concerns have emerged in areas such as mental health5. Lång therefore argues that safety should be demonstrated before potential clinical benefits of medical AI are considered, particularly in screening programs in which any harms would have an impact on millions of people.

Likewise, AI performance may deteriorate over time. Shah et al. found that model iterations and changes in electronic record documentation could lead to performance drift, after deployment of their ChatEHR AI system. Addressing these concerns in practice would require ongoing oversight after AI systems are deployed, supported by governance structures that allocate clear responsibility to healthcare organizations and clinicians for monitoring performance, responding to emerging risks and implementing corrective actions.

The long-term success of clinical AI adoption also depends on whether healthcare staff are empowered to use these technologies in routine practice and perceive their clinical value. This challenge was evident in a pilot evaluation of an AI-based clinical decision support system, in which adoption declined over 4 weeks as clinicians became less engaged with the tool.

Schaekermann et al. therefore emphasize the need for AI tools to earn the trust and acceptance of clinicians and patients. This will require prospective real-world studies that demonstrate how these AI technologies affect clinical workflows and the day-to-day experiences of healthcare staff. Patients’ perspectives are also key, as experience from fields such as mental health have shown that people with lived experience can identify barriers and unintended consequences of digital technologies that are often overlooked by the assumptions of developers and researchers6.

Making AI work in healthcare will ultimately require a shift in how these technologies are evaluated — before, during and after deployment. In addition to benchmark performance, greater emphasis is needed on how AI technologies affect clinical outcomes, care delivery, cost-effectiveness and health service performance, supported by emerging AI evaluation frameworks7. Answering these questions would mean considering implementation from the earliest stages of research and using prospective study designs such as hybrid effectiveness–implementation clinical trials and quasi-experimental studies. For example, the USA has recently funded $62.7 million for the ADVOCATE program, which aims to deploy autonomous AI devices for cardiovascular care and address the prospective evidence gap surrounding their real-world use and impact.

As health systems and policymakers2 move to accelerate the deployment of AI in healthcare, stronger evidence is needed on whether these technologies are having a meaningful impact. The experiences described across in this issue show that successful integration of AI technologies into health systems will depend on how effectively they work alongside the people expected to use them. Ultimately, the value of medical AI will depend on whether it helps deliver better care, not just better performance.

来源:Nature Medicine · nature.com