The FDA wants companies to be more transparent about their AI algorithms. | Gerard Julien/AFP via Getty Images
This week, the Food and Drug Administration issued guidance documents concerning artificial intelligence in medical devices and drug discovery that highlight its focus on algorithm transparency. The documents offer nonbinding advice to the drug and medical device industries on technological advances.
Drug development: One draft guidance puts guardrails around AI models used to help create or support data submitted to prove a drug’s safety, effectiveness and quality.
The agency wants applicants to demystify their algorithms and provide specific details about what the models do, how they’re supposed to be used, how risky they are to patients, how well they perform and how companies plan to monitor them. It also asks companies to design and complete credibility assessments for their models.
The guidance builds on a 2023 discussion paper the agency published on the use of AI in drug discovery and development. It received more than 800 comments from 65 organizations, POLITICO has reported. Between 2016 and 2022, the FDA reviewed more than 500 drug submissions in which AI was used. In 2021 alone, there were 132 such applications.
Medical devices: A second draft guidance focuses on the design and development of AI in medical products that undergo the medical device submission process. It recommends how companies can share their data to ensure their products are unbiased, safe and effective. It also reveals how the agency plans to conduct postmarket surveillance, calling on companies to submit a performance monitoring plan, which may be a condition of approval for certain devices, particularly those classified as novel or de novo.
The agency has published AI guidance since 2021, largely focusing on medical devices and clinical decision support. In October, the agency published a discussion piece in JAMA expressing a need for “flexible mechanisms” to keep up with AI advancements in pharma and health care.
What’s next: A major ongoing debate focuses on how to monitor health care AI after its release into the world. FDA Commissioner Robert Califf has said the agency needs more staff to adequately review all health AI. A report from the Government Accountability Office suggests the FDA might need new authority to tackle postmarket AI surveillance.
For now, the guidance suggests the FDA will rely on company-produced monitoring plans. But experts in the field suggest regulators might need to consider an entirely new paradigm ensuring AI is safe after deployment.
WELCOME TO FUTURE PULSE
Chappaquiddick, Mass. | Richard Schumaker
This is where we explore the ideas and innovators shaping health care.
Many women delay trips to the gynecologist because they're uncomfortable with cervical cancer screenings. | Getty Images
A new survey shows that over half of women delayed cervical cancer screenings because they were uncomfortable with the process.
Four thousand women die of cervical cancer every year, and a significant portion of them don’t get screened. A 2021 Centers for Disease Control and Prevention study found that among 376 women who survived invasive cervical cancer, 60 percent didn't receive the recommended cervical cancer screening using a Pap test in the five years prior to their diagnosis.
So why aren’t women getting screened? More than half of women in a new survey said they delay trips to the gynecologist because of fear or general discomfort with the process, according to the survey. This was true for women in the U.S., U.K., Canada and Indonesia. Roughly half had difficulty finding time to go get a screening.
The online survey included responses from more than 2,500 women older than 18 in the U.S., the U.K., Canada and Indonesia. It was conducted by Harris Poll on behalf of Becton Dickinson and Company, a New Jersey-based diagnostic test manufacturer. BD has developed a self-sampling human papillomavirus, or HPV, test that the Food and Drug Administration recently approved for use at doctors’ offices.
But the real potential lies in expanding access to screenings by making these tests available for home use.
A 2024 study published in eClinicalMedicine found that when U.K. primary care doctors offered take-home HPV tests to women, monthly screening rates increased by 22 percent among patients who were overdue.
What’s next: The National Cancer Institute is conducting a nationwide clinical trial across 25 clinical sites to evaluate various at-home tests for HPV. The program, called the Last Mile Initiative, kicked off last summer.
The program hopes to generate the data needed to obtain FDA approval of home tests for cervical cancer.
DANGER ZONE
An FDA-approved genetic test for opioid addiction risk is no more accurate than a coin toss. | Patrick Sison/AP
Researchers at the University of Pennsylvania’s Perelman School of Medicine analyzed a diverse sample of nearly half a million people exposed to opioids, out of which nearly 34,000 had an opioid use disorder. The results showed that the genetic variants the researchers say underpin the test were accurate in about 53 percent of the cases.
“It’s no better than chance,” Dr. Henry Kranzler, a psychiatry professor and the director of the Center for Studies of Addiction at the school, who led the study, told Carmen.
How it works: The AvertD by AutoGenomics test uses 15 genetic variants and machine learning to determine whether a patient who has never been prescribed opioids is at risk of developing addiction if they get opioid pain medication for up to 30 days to treat acute pain.
The information from the test shouldn’t be used alone but as part of a complete clinical evaluation and risk assessment, the FDA said when it approved the test a year ago.
But Kranzler said tens of thousands of genetic variants in a person’s genome could contribute to addiction, and the 15 the test focuses on aren’t sufficient to predict someone’s opioid addiction risk.
Why it matters: The test could harm patients by hampering access to pain medication for those at low risk for addiction or giving a false sense of security to those at high risk, Kranzler said.
Clinicians could better predict opioid use disorder risk using an individual’s age and sex than the 15 genetic variants, Kranzler and his colleagues wrote in the study.
He also pointed to a questionnaire that can be used successfully to predict the risk in many patients.
“The whole issue of genetics and addiction is really understudied, it’s underfunded by the [National Institutes of Health], it’s not considered as high a priority as it needs to be,” he said.
The FDA and AutoGenomics didn’t respond to a request for comment.