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.
July 25, 2025
In this week’s Duke AI Health Friday Roundup: how patients feel about doctors using AI; drug-eluting composites for fixing bones; the value of publishing null results; why thinking more makes some reasoning models perform worse; flossing tested as alternative to needles for vaccination; study implicates increased calories, not lack of activity, in obesity rates; LLM helps adapt radiology reports for patients; much more:
AI, STATISTICS & DATA SCIENCE
- “In line with prior research, our results indicate that the public has certain reservations about the integration of AI in health care. While the present effect sizes are relatively small, in particular regarding AI use for administrative purposes, they may be highly relevant as trust in health care practitioners is closely linked to subjective treatment outcomes.Potential reasons for existing skepticism may include concerns that physicians rely too much on AI and that the use of AI could reduce patient-physician interactions as well as concerns about data protection and rising health care costs.” A research letter by Reis and colleagues published in JAMA Network Open presents results from a survey that gathered information about patients’ attitudes about physicians who make use of AI tools.
- “These findings suggest that while test-time compute scaling remains promising for improving model capabilities, it may inadvertently reinforce problematic reasoning patterns. Our results demonstrate the importance of evaluating models across diverse reasoning lengths to identify and address these failure modes in LRMs.” A research paper by a group of Anthropic AI authors, available as a preprint from arXiv, probes the possible explanations for why some large reasoning model AIs seem to perform worse, the longer they spend mulling over a problem.
- “To develop RadGPT, the Stanford team took 30 sample radiology reports and extracted five concepts from each report. With those 150 concepts, they developed explanations for them and three question-and-answer pairs that patients might commonly ask. Five radiologists who reviewed these explanations determined that the system is unlikely to produce hallucinations or other harmful explanations…the current RadGPT model depends on a human radiologist dictating a report, and only then will the system extract concepts from what they have written.” A web article by Stanford’s Human-Centered Artificial Intelligence Institute describes the creation of “RadGPT,” a customized large language model, still currently in development, designed to extract information from highly technical radiology reports and adapt it for patients.
- “Compared to traditional statistical and ML models, LLMs exhibit superior generalizability and can provide dynamic and longitudinal assessment of aging. Moreover, interpretability analyses suggested that LLMs were capable of reliably assessing individual aging. Our research provides evidence for the application of LLMs in aging assessment for large-scale general populations.” In a research article published in Nature Medicine, Li and colleagues describe the creation and evaluation of a large language model designed to estimate biological age from information gathered in routine health examinations.
BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH
- “Comparisons of energy expenditure across populations strongly suggest that increased energy intake (i.e., caloric consumption and absorption) is the primary factor promoting overweight and obesity with economic development. This view is supported by measures of total expenditure and weight change, which together provide an estimate of energy intake… Our analyses suggest that increased energy intake has been roughly 10 times more important than declining TEE in driving the modern obesity crisis.” A research paper published in PNAS by McGrosky and colleagues presents an exhaustive analysis that sampled a diverse array of global populations and found that increased caloric intake, not levels of physical activity, are the primary driver of obesity.
- “Flossing may be good for more than getting your dentist off your back—one day, it may also protect you from the flu. In an unorthodox approach to needle-free vaccines, researchers have developed a special kind of floss that can deliver proteins and inactive viruses to mice’s gumlines and trigger immune responses that protect against infectious disease…” A news article by Science’s Annika Inampudi reports on recent work that has yielded a new method for vaccination that avoid the use of needles.
- “Our findings show that USAID-supported efforts have helped to prevent more than 91 million deaths across all age groups, including 30 million deaths among children. High levels of USAID funding were associated with a 15% reduction in all- age and all-cause mortality, a 65% reduction in mortality from HIV/AIDS, a 51% reduction from malaria, and a 50% reduction from neglected tropical disease.” An analysis published in the Lancet by Cavalcanti and colleagues shows the global impact of USAID funding on mortality and provides a forecast for the consequences of recent cuts in funding to the organization.
- “…this proof-of-concept study establishes a robust foundation for utilizing modified dental resins in the fabrication of multifunctional, drug-eluting orthopedic implants. The convergence of 3D printing technology with antimicrobial-eluting capabilities represents a significant advancement toward personalized, infection-resistant bone fixation devices for high-risk orthopedic interventions.” An article published in Scientific Reports by Zhang and colleagues describes the initial testing of 3-D printed bone nails for orthopedic surgery that elute antibiotic drugs to help resist infection.
COMMUNICATION, Health Equity & Policy
- “Analysis of pre- and post-trial data from 2,896 employees across 141 organizations in Australia, Canada, Ireland, New Zealand, theUK and the USA shows improvements in burnout, job satisfaction, mental health and physical health—a pattern not observed in 12 control companies. Both company-level and individual-level reductions in hours are correlated with well-being gains, with larger individual-level (but not company-level) reductions associated with greater improvements in well-being. A research article published by Fan and colleagues in Nature Human Behavior finds a number of benefits following from a four-day work week.
- “Some 1,489 respondents had generated null results and agreed that they are important to share but had not yet done so. Most of that group expressed concerns about publishing them: 69% didn’t think null results would be accepted for publication; 52% didn’t know which journals would consider publishing research with null results; 19% worried that their institution or funder wouldn’t cover publishing costs; and 21% were concerned that they’d be viewed negatively by their peers.” Nature’s Laurie Eudesky reports on a Springer Nature survey that assessed scientists’ attitudes toward publication of null results – and the difficulties they encounter in getting them accepted for publication.
- Given the high decisional uncertainty during pregnancy about vaccinating children after birth, there may be value in intervening during pregnancy to proactively support families with childhood vaccination decisions. Strengths of our study include the national scope of the surveys….Future studies with longitudinal follow-up may shed further light on evolution of vaccination decisions from pregnancy to parenthood, and the effectiveness of intervening proactively during pregnancy. A research letter published by Vasudevan and colleagues in JAMA Network Open describes findings from a pair of nationwide surveys that probed intentions with regard to infant vaccination before birth and afterward.
- “…rapid data center growth is the foundation of an industrial strategy that is betting on AI innovation and energy dominance as the keys to US technological supremacy. Yet leveraging extractive AI infrastructure for geopolitical advantage sets up a tension that may well prove unresolvable: achieving computing abundance requires exhausting the planetary bounds of our ecological resources, energy, and labor, all of which are required to support this industrial development.” A Data & Society policy brief by Kneese and Woluchem offers a critical examination of ramped-up growth in data centers meant to serve the needs of the burgeoning AI sector.
