EPtalk by Dr. Jayne 3/20/25

Google Chief Health Officer, Karen DeSalvo, MD, MSc blogged about six health AI updates that were recently covered at the company’s annual The Check Up event. Top on the list was “helpful health results in Search.” She notes that since the debut of AI Overviews, users are “asking longer, more complex questions” and states a goal that health-related overviews will “continue to meet a high bar for clinical factuality.”

It still surprises me when I hear my physician colleagues using Google to look up medical information compared to using a more validated healthcare-specific resource. I certainly wouldn’t want to be on a witness stand explaining where I got my information, as opposed to a peer-reviewed journal article or a national consensus guideline.

She also notes the release of medical records APIs in Health Connect, allowing systems to leverage core medical elements such as allergies, medications, immunizations, and lab results via FHIR. Another item highlighted is the company’s FDA clearance for the Loss of Pulse Detection features on the Pixel Watch 3. It can generate a call to emergency medical services if the wearer is unresponsive. It’s apparently been available in the EU and will roll to the US later this month.

Fourth on the list is an “AI co-scientist” that is intended to “help biomedical researchers create novel hypotheses and research plans” after combing through the scientific literature. Development partners include Imperial College London, Houston Methodist, and Stanford University. Not included in the writeup is the Oxford comma, which I have compulsively added to the list.

Fifth is TxGemma, which aims to speed AI-enabled drug discovery. The solution can manage text as well as molecular structures. I have zero experience with that technology, but it brought back not-so-fond memories of building hydrocarbons during organic chemistry, which represents eight college credit hours of my life that I will never get back.

Last on the list is a tool called Capricorn, which is designed to support the rapid identification of personalized cancer treatments through the integration of public medical data and de-identified patient data. It is supposed to synthesize the literature along with potential treatment options. It is being developed in partnership with the Princess Maxima Center for pediatric oncology in the Netherlands. It will be interesting to revisit these solutions in six, nine, or 12 months to see which of them show real promise and whether any of them have fizzled.

Another Google feature that wasn’t covered in the story and about which I am less than enthusiastic is the “What People Suggest” search feature that will allow patients to “search through online commentary from patients with similar diagnoses.” Patient experience is certainly important, and patients can be powerful advocates and support systems for each other. However, just because another patient with a similar condition was treated in one way doesn’t mean it’s the right treatment for another patient.

It’s a rare patient care day that I don’t see information patients have researched on the internet or from patient forums, and a good chunk of it is irrelevant to the patient in front of me. It takes a great deal of time to have these discussions with patients and quickly becomes untenable for primary care physicians who are carrying panels of thousands of patients. Patients become frustrated when they learn that treatments advocated by others may not be standard of care or in fact might be harmful. I wish we could spend some public health dollars helping patients learn how to better analyze the information they see on social media and the internet, but we all know there aren’t enough public health dollars as it is.

A recent article in Science reviews an AI tool that can evaluate blood samples and determine the likelihood of infections, autoimmune diseases, or the response to a vaccine. It looks at the genes that code for B and T immune cell receptors and was able to identify patients with COVID-19, HIV, type 1 diabetes, lupus, those recently vaccinated, and those who met none of those criteria.

I found immunology to be one of the most interesting topics in my medical school curriculum, although I struggled with it due to a professor who really didn’t want to teach students and made it clear he preferred to be in the lab. Discoveries like this might just make me want to learn more about it again.

Good news from a payer (for once): Optum Rx, which is part of UnitedHealth Group, has announced its intention to update prior authorization requirements for 80 prescription drugs. Although several news articles about the announcement used the phrase “remove prior authorization requirements,” it’s not exactly what it sounds like. From what I understand, the modifications planned will impact “reauthorizations,” which is where a physician has to obtain approval to continue a drug that a patient is already taking. This is explained in the press release with examples.

A “necessary” reauthorization might occur for “drugs that have safety concerns, need ongoing monitoring for dose adjustments, require additional tests, or may have alternative therapy considerations.” Those that will be reduced are for drugs where “there is minimal additional value in reauthorizing an effective, lifelong treatment.” They expect a 25% reduction in reauthorizations. No list of drugs was provided, so I wonder if they haven’t fully identified the list yet or whether they’re keeping it to themselves in hopes that some requests will experience attrition during the process because physicians are simply exhausted.

I enjoyed reading a recent commentary by NYU Grossman School of Medicine ethicist Art Caplan, PhD. He was reacting to Elon Musk’s request that patients upload copies of their medical imaging studies to help train his Grok AI solution. I often read Caplan’s editorial pieces and respect his straightforward take on issues. He notes that AI hallucinations are real, and “If you go out and take random information submitted by a subpopulation of people, not representative of everybody, you’re going to get many false findings.”

He goes further to discuss the perils of not knowing the attributes of a particular image, such as whether it’s accurate, the demographic characteristics of the patient, and more as far as being able to have training data where bias is mitigated. He also notes that there are no assurances of privacy for any images that are sent.

My favorite quote from his comments is this: “The last big issue is, why should we be doing this for free? Elon Musk is a gazillionaire. If he wants information, why doesn’t he go out and pay a representative sample of people to undergo tests, establish what a normal baseline looks like, and then try to explore what disease baselines look like? That’s what we need to have good automated technology to help diagnosis — and note that I said help it, not replace it. If there’s no baseline and people are just randomly firing in medical tests, you’re not going to have an accurate AI diagnostician; you’re going to have a mess.” Thanks for telling it like it is, Dr. Caplan.

What do you think of the idea of crowdsourcing medical images for an AI training dataset? Leave a comment or email me.

Email Dr. Jayne.