AI Health
Friday Roundup
The AI Health Friday Roundup highlights the week’s news and publications related to artificial intelligence, data science, public health, and clinical research.
September 12, 2025
In this week’s Duke AI Health Friday Roundup: LLMs extract safety data from product labels; daylight saving time maybe not that bad; the limits of vibe coding; getting a handle on Alzheimer’s heritability; AI impersonates docs to pitch dodgy products; calculating ROI for COVID vaccination; Medicare Advantage plans might be contributing to hospital discharge bottlenecks; much more:
AI, STATISTICS & DATA SCIENCE
- “…we found generative LLMs, particularly GPT-4, was [sic] effective in extracting adverse reactions (ARs) and drug interactions (DIs) with no additional training or fine tuning. The results show that for these extraction tasks, performance was not sensitive to different prompting strategies. Furthermore, we found that incorporating shot prompting did not improve performance.” A research article published in the journal Drug Safety by Gisladottir and colleagues report on the use of large language models to extract key drug safety information from product labels.
- “By skipping that entire process and going straight to the artifact, vibe coding now allows any individual to borrow that credibility, without actually earning it….in the system of output-focused short feedback loops that we have constructed to measure the software development lifecycle, this is “close enough.” It resembles productivity, but it is not actually making us more productive. The lines of code increase (as does the time it takes to QA!), but upon closer investigation those LOCs either do nothing or actually cause harm…The result of tools that put solutioning within our reach is that we are disincentivized from taking the time to construct the problem.” An essay by Pavel Samsonov at The Product Picnic explores some of the pitfalls of reliance on AI-assisted “vibe coding” for software development.
- “The report comes from the Census Bureau and shows that the rate of AI adoption by large companies—that is, firms with over 250 employees—has been declining slightly in recent weeks. The report is based on a biweekly survey, dubbed Business Trends and Outlook (or BTOS), of some 1.2 million U.S. firms. The survey, which asks businesses about their use of AI tools, such as machine learning and agents, found that—between June and now—the rate of adoption had declined from 14 to 12 percent.” Gizmodo’s Lucas Ropek reports on recently published Census Bureau data that shows a dip in the rate of AI adoption by larger US businesses.
- “The posts are part of a global surge of frauds hijacking the online personas of prominent medical professionals to sell unproven health products or simply to swindle gullible customers.…While health care has long attracted quackery, A.I. tools developed by Big Tech are enabling the people behind these impersonations to reach millions online — and to profit from them. The result is seeding disinformation, undermining trust in the profession and potentially endangering patients.” A report by the New York Times’ Steven Lee Myers, Alice Callahan, and Teddy Rosenbluth documents widespread use of AI tools to impersonate well-known physicians in order to hawk various scam treatments.
BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH
- “In a sweeping analysis of nearly 170,000 patients over a decade, researchers at Duke University School of Medicine found no significant increase in heart attacks during the weeks surrounding DST transitions — in neither spring nor fall….The findings published Sept. 9 in JAMA Network Open challenge earlier, smaller studies that suggested the spring time change, which robs people of an hour of sleep, could lead to a rise in cardiovascular events.” A web article by the Duke University School of Medicine’s Shantell Kirkendoll highlights research by Duke investigators that suggests the ostensible health effects of Daylight Saving Time may not be as bad as previously thought.
- “…we identified LOAD heritability estimates from 6 twin studies and 17 genome-wide association studies, all conducted in populations of European ancestry. We demonstrate that these heritability estimates for LOAD vary considerably. The variation reflects not only differences in study design and methodological approaches but also the underlying study population characteristics. Our findings indicate that commonly cited heritability estimates, often treated as universal values, should be interpreted within specific population contexts and methodological frameworks.” A systematic review published in PLOS Genetics by Liu and colleagues finds substantial variation in estimates of heritability for Alzheimer disease across different studies.
- “In identifying and avoiding the shortfalls that have plagued other research in the field, the new papers make a ‘huge contribution to the scientific literature,’ says Matthew Meyerson, a geneticist at Dana-Farber Cancer Institute and Harvard Medical School who was not involved in either study. The studies’ ‘highly accurate’ picture of the species found alongside tumors will help scientists learn more about microbes that might cause or respond to a person’s cancer, he adds.” Science’s Catherine Offord reports on a pair of studies published in Science Translational Medicine (one by Ge and colleagues, one by Gihawi and colleagues) that revisit flaws found in previous studies of cancer microbiome sequencing and blood tests developed based on those studies, while also breaking some new ground.
COMMUNICATION, Health Equity & Policy
- “A key government study about alcohol and its health harms will not be released publicly, despite several years of taxpayer-funded work and a growing body of evidence connecting drinking with disease…. A Department of Health and Human Services spokesperson said the alcohol study had been provided to HHS and the Department of Agriculture for consideration, but declined to answer questions about whether the final report would be released or used to develop the dietary guidelines.” STAT News’ Isabella Cueto reports on news that a long-awaited Substance Abuse and Mental Health Services Administration (SAMHSA) report on the health risks of alcohol consumption will not be released.
- “Globally, for every dollar invested in vaccine development and delivery, the estimated returns ranged from $42.0 to $775. To our knowledge, this study provides the first economic evidence that the unprecedented, rapid investments made into COVID-19 vaccines paid off both in lives saved and long-term economic benefits. The overall findings were robust to changes in the approaches used (welfarist or extra-welfarist, discounted or undiscounted) and in various sensitivity analyses.” An analysis by Benjamin and colleagues, available as a preprint from medRxiv, evaluates return on investment from the first year of COVID vaccination.
- “A researcher recalled wondering about the level of six liquid nitrogen tanks used to maintain samples ‘going back to the 80s’ and checked on them to find that ‘all of them were bone dry.’ At first, the researcher worried that ‘with everything going on,’ they may have neglected them….Then they checked with their business office and learned that ‘our lab hadn’t purchased any liquid nitrogen since 2016.’” A difficult-to-assimilate report at Retraction Watch lays out an astonishing account of fakery and data manipulation by a laboratory research assistant.
- “…from 2017 to 2023, Medicare Advantage beneficiaries experienced increasing rates of extended hospital stays relative to traditional Medicare beneficiaries, even after robust adjustment for changes in case mix. Medicare Advantage beneficiaries disproportionately drove rising rates of extended hospital stays in the whole Medicare program over the study period even though total admission volume fell. The difference in patterns of hospital use was magnified by the rapid growth of the Medicare Advantage program over this period.” An analysis of hospital discharge delays published in JAMA Internal Medicine by McGarry and colleagues finds that requirements related to Medicare Advantage insurance plans may be contributing to those delays.
