FDA 批准的人工智能设备中近一半并非基于真实患者数据

FDA 批准的人工智能设备中近一半并非基于真实患者数据
Nearly Half of FDA-Approved AI Devices Not Based On Real Patient Data

原始链接: https://www.zerohedge.com/medical/nearly-half-fda-approved-ai-devices-not-based-real-patient-data

北卡罗来纳大学的研究人员表示，人工智能 (AI) 医疗设备需要更严格的测试程序。这一结论是在对美国食品和药物管理局 (FDA) 自 1995 年以来的批准进行广泛审查后得出的。值得注意的是，这些人工智能医疗设备中几乎一半没有使用实际患者数据进行测试。相反，一些人使用模拟或虚构的图像，导致人们对其在治疗真实患者时的可靠性和有效性产生疑问。尽管有许多优点，例如在放射学检查中检测癌症或中风等疾病，但由于这些设备在真实患者身上的表现证据不足，人们仍然担心与这些设备相关的风险。研究人员敦促 FDA 强制要求进行临床验证（对活体受试者进行测试），以确保设备有效运行。目前，只有不到十分之六的人工智能医疗设备满足这一要求。经过彻底评估后，研究人员建议创建“黄金标准指标”来评估人工智能医疗设备的安全性和效率。放射应用占研究设备的大部分，占案例的 75%。这些设备通常属于中等风险 II 类，包括各种诊断工具、手术器械以及起搏器和助听器等疗法。为了提高公众接受度并建立对该技术的信心，研究人员强调发布授权设备的大量临床验证数据。研究人员表示，增加临床验证将有助于实施并培养公众信任。首席研究员 Sammy Chouffani El Fassi 承认该领域仍处于起步阶段，但指出，经过审查的人工智能医疗设备风险很小。它们并不是为了取代医生而设计的；相反，他们的目标是补充和简化医疗保健专家的任务。然而，她指出，需要进一步调查，以揭示这些设备的广泛使用所造成的任何长期后果。该研究凸显了人工智能医疗设备的激增，FDA 每年的批准数量从 2016 年的仅有 2 个跃升至 2022 年的 69 个。为了征求意见，我们向 FDA 和两家相关设备制造商发送了请求，但都没有回应。

原文

Authored by Huey Freeman via The Epoch Times (emphasis ours),

Researchers from the University of North Carolina have called for more rigorous testing of artificial intelligence (AI)-powered medical devices, following a comprehensive study of nearly three decades of U.S. Food and Drug Administration (FDA) authorizations.

The study, published in Nature Medicine, found that nearly half of AI medical devices authorized by the FDA were not based on real patient data, raising concerns about their safety and effectiveness in real-world patient care.

Some devices used simulated images, not real patient data, which technically didn’t qualify as testing in real patients, also known as clinical validation.

Although AI medical devices serve many useful purposes, including detection of cancer and strokes on radiology scans, this study shows they also bring with them potential dangers.

“We shared our findings with directors at the FDA who oversee medical device regulation, and we expect our work will inform their regulatory decision making,” Sammy Chouffani El Fassi, a doctor of medicine candidate at the University of North Carolina Medical School and first author, said in an interview with The Epoch Times.

The study, completed in about 18 months, included eight authors, as well as a large team of consultants from academic institutions and corporations.

The study highlighted the rapid growth of AI medical devices, with FDA authorizations increasing from two to 69 annually between 2016 and 2022.

A Need for Higher Standards

Researchers recommend the FDA require clinical validation for all devices, meaning testing on real patients so scientists can see that they work, Chouffani El Fassi said.

Their analysis revealed that only 56 percent of approved devices had this validation.

After analyzing FDA authorizations from 1995 to 2022, researchers recommended establishing a “gold-standard indicator” of safety and effectiveness. Most authorized devices were for radiology, with 75 percent in this category. Nearly all were classified as intermediate-risk class II devices. Class II devices include diagnostic devices like X-ray machines, surgical devices like catheters and sutures, and therapeutic devices such as pacemakers and hearing aids.

“For the public to accept FDA authorization as an indication of effectiveness, the agency and device manufacturers must publish ample clinical validation data,” the researchers wrote.

Effectiveness Proven By Testing in Patients

“We believe having more clinical validations published will reduce barriers to implementation,” Chouffani El Fassi, said. "It will increase the public trust in the whole technology. It is powerful what this technology can do. It can predict the onset of disease before it even starts.”

Chouffani El Fassi acknowledged that the field is relatively new and that the full extent of potential harm is unknown.

The devices analyzed in the study were categorized as low-risk, he said. They are not intended to replace doctors but rather to assist and augment their work.

“There is a limit to what kind of harm they can do to people,” Chouffani El Fassi said. “That is why they get authorized. At the end of the day even if an AI health tool helps read a chest X-ray, for example, a human physician is going to read over that X-ray. The AI helps triage the scans and helps the physicians to look over some scans sooner.”

Testing in Patients Can Be Simple

Testing in patients is not always performed because they are a rigorous, costly process, Chouffani El Fassi said.

The study sought to establish a standard for clinical validation.

Researchers prefer prospective validation as they provide stronger evidence, according to Chouffani El Fassi. Prospective validation tests the AI machine on new data while retrospective validation tests the AI machine on historical data. In prospective validation tests, researchers may conduct randomized controlled trials to compare device users to a control group.

“That is the gold standard for medicine because you are comparing the group of health care professionals that used the device and a control group that did not use the device,” Chouffani El Fassi said.

The Epoch Times reached out to the FDA for comments. Two device manufacturers declined to comment.