You’re reading the web edition of STAT’s Health Tech newsletter, our guide to how technology is transforming the life sciences. Sign up to get it delivered in your inbox every Tuesday and Thursday.
Pfizer recently rolled out a new generative AI product, called Health Answers by Pfizer, designed to provide consumers trustworthy information about common questions. After punching in a query, like “how can I better treat my asthma,” the product hums away, much like ChatGPT or similar, before returning a response derived from “verified sources,” including a citation. It feels a little bit like typing queries into Google and getting answers from WebMD or another SEO-optimized health resource, without having to actually read (or skim) a whole article.
People spend a lot of time Googling health questions, so here’s a faster way to do that, in theory. Pfizer warns that responses are not medical advice, but there’s always a risk that people will take responses as such. The risk of bad information reaching users may be higher when an LLM is generating responses, as opposed to a human writing edited content.
A few notes after using Health Answers by Pfizer. Plus, a few responses from chatting with a Pfizer spokesperson over email.
I asked about my knee that hurts. Health Answers gave me all kinds of plausible answers and suggested that I should probably go see a doctor.
Health Answers once warned me that a response was “unverified.” It said, “we did not have enough content from our trusted sources to generate a verified answer using Health Answers by Pfizer’s standards.” It did give me an answer anyway.
I asked, “can you die from touching fentanyl?” Health Answers warned me that “this could be serious” and responded: “Yes, touching fentanyl can be very dangerous and potentially life-threatening.” ChatGPT’s response? “No, you cannot die from simply touching fentanyl. This is a common misconception.” (Admittedly, this is a useful question only if you are trying to break an LLM. After I told Pfizer about it, they updated the product.)
Can Pfizer make money from this? Pfizer says that Health Answers is independent from its pharmaceutical business. The company also runs PfizerForAll, a direct to consumer portal that offers to connect users to telehealth providers that can prescribe Pfizer drugs or whatever treatment is appropriate. There are links to Health Answers on PfizerForAll, and Health Answers provides links to find a doctor on services like BetterHelp, Sesame, Zocdoc, and of course, PfizerForAll’s telehealth service which is run by UpScript. The Health Answers website warns that “when you click on or make a purchase or an appointment through links on our site, we may earn a fee.”
What about data? It’s not required, but Pfizer prompts users to create a profile, including information about any health conditions, which will be used to provide an “enhanced and personalized experience.” As part of this process, you can also opt-in to receive marketing from “Pfizer Affiliate Partners that are personalized” based on your use of the site, including the health information you have shared. Pfizer warns that “certain information, such as health information, may be considered sensitive.”
What tech is under the hood? In a LinkedIn post, Pfizer executive Dennis Hancock thanked Microsoft for helping Pfizer with the product. Microsoft has close ties to OpenAI. Pfizer told me it used “a public LLM with proprietary enhancements to optimize answers to health and wellness questions.”
What is being done to make sure nothing bad happens? Pfizer said: “Safeguards include oversight from an independent scientific and technology advisory board, filters to help ensure answers are faithful to the source material, clinician audits for accuracy, and clear disclaimers throughout the platform on the use of AI.”
HHS leaders to miss HIMSS. Possible cuts loom
The HIMSS conference is underway in Las Vegas, but dozens of federal officials who were scheduled to attend aren’t there, owing to a pause in event travel instituted by the Trump administration. As STAT’s Brittany Trang reported, at least 46 officials from health department agencies like FDA, CMS, CDC, and ASTP/ONC were originally scheduled to attend the most prominent health IT conference of the year. We confirmed at least 28 federal officials were skipping the event.
In the past, HIMSS has been a key moment for regulators to update industry on pending plans. For example, this year, ASTP, which oversees electronic health record software, data interoperability, and other health IT issues, was scheduled to present on a regulation called HTI-1, which mandated transparency about how some health AI products work, and about the future of TEFCA, an initiative that’s meant to streamline sharing of data between health organizations.
Speaking of ASTP, layoffs targeting probationary workers hit the organization a few weeks ago, and there are swirling rumors about its fate under a Trump administration intent on downsizing the government. ASTP, which stands for the Assistant Secretary for Technology Policy, is less than a year old and was created as an expansion of the role of Micky Tripathi, who led the Office of the National Coordinator for Health Information Technology, and ASTP, until January. Part of the goal was to help better oversee the federal health department’s use of artificial intelligence.
This expanded office is in flux. Alicia Rouault, who was appointed as chief technology officer at ASTP in January, updated her LinkedIn profile to say she no longer works at HHS. Longtime ONC staffer Steven Posnack is serving as acting head of the organization. With the Trump administration’s new directive calling for agency leaders to find further ways to trim staffing, it’s possible ASTP/ONC could be targeted for additional cuts soon.
STAT is dedicated to following the tremendous changes underway at HHS. If you know something, consider reaching out to a STAT reporter. My email is [email protected] and you can reach me on Signal at mariojoze.13.
DOJ probes device maker whose tests lead to lucrative diagnoses for UnitedHealth
Casey Ross writes: STAT exclusively reported on Monday that the Department of Justice is investigating Semler Scientific for possible violations of the False Claims Act. Semler’s only product, QuantaFlo, is used to test patients for peripheral artery disease. STAT reported last year that QuantaFlo, which uses a proprietary algorithm, generated a flood of questionable, and lucrative, diagnoses for UnitedHealth and other insurers. Each diagnosis is worth about $3,000 a year in reimbursement from Medicare, which pays insurers more to cover sicker patients. UnitedHealth did not respond to a request for comment from STAT. Semler also did not respond. Semler disclosed the DOJ probe in an SEC filing that indicated settlement discussions with the government began and ended on Feb. 11. Company officials stated in the filing that they would vigorously defend themselves against any charges.
How AI can warp incentives in health care
Years ago Ken Mandl coined the term “biomarkup” for how commercial interests could influence the development and use of seemingly objective measures of medical status, like laboratory tests. For example, a drugmaker who might stand to gain from increased testing that’s likely to increase diagnoses might push providers to use it more.
In an interview with STAT’s Katie Palmer, Mandl discusses how the concept might be applied in the world of artificial intelligence. For example, AI-based revenue cycle tools might push clinicians towards more profitable care pathways. “Wherever there are financial flows, there are going to be parties that are interested in optimizing those financial flows,” said Mandl, who is director of the computational health informatics program at Boston Children’s Hospital.
In some cases AI products serve a similar function as lab tests, and thus might be gamed the same way.
“Many, many, many algorithms are, in a sense, biomarkers — that will drive a diagnosis and therefore treatment, or to help select a treatment, or to put someone into a risk group that may need preventive treatment, or that may drive someone to get a referral,” he said. “If you understand this idea of biomarkup, you can see that it could potentially become extremely pervasive in the AI era.”