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.
March 21, 2025
In this week’s Duke AI Health Friday Roundup: articles with three-part titles tend to get cited more; using AI to hunt errors in scientific literature; backpropagation for general-purpose LLMs; analogs of addiction in internet use; Faraday’s notebooks unveiled to public; journal click-through suffers as AI-generated summaries proliferate; framework allows selection of best AI tools for task; LLMs in wide use for search and research return substantial proportion of errors; much more:
AI, STATISTICS & DATA SCIENCE
- “The projects have tentative support from academic sleuths who work in research integrity. But there are also concerns over the potential risks, including that the tools could be used maliciously and before they are ready. How well the tools can spot mistakes, and whether their claims have been verified, must be made clear, says Michèle Nuijten, a researcher in metascience at Tilburg University in the Netherlands.” Nature’s Elizabeth Gibney reports on the emerging use of large language models to police published research papers for errors – and the potential pitfalls of such an approach.
- “Suppose we could speed up the unlearning of what is wrong but widely ‘known’ through advancing new distinctions between noise and accurate observations, translating hierarchies of abstraction in structured vocabularies to new worldviews, and improving the integration of both old and new facts into new knowledge. All of this would be widely impractical, except for the advances in AI that can now drive the re-writing of science in the setting of a scientific revolution.” A perspective article published in the Journal of the American Medical Informatics Association by Leslie Andrew Lenert explores future directions for the National Library of Medicine as the institute adapts to the advent of AI technologies.
- “Recently, a team of Stanford researchers has devised a new way to evaluate how well out-of-the-box language models perform routine health care tasks. That includes coordinating care, coming up with treatment plans, documenting visits, and more — not just answering multiple-choice questions. The results of the effort, called MedHELM, has already produced a chart that can help health systems and AI developers choose language models that are best suited for different medical tasks.” STAT News’ Brittany Trang profiles Stanford’s MedHELM system for evaluating whether a given language model is fit for purpose when it comes to healthcare applications.
- “A new study from Columbia Journalism Review’s Tow Center for Digital Journalism finds serious accuracy issues with generative AI models used for news searches. The research tested eight AI-driven search tools equipped with live search functionality and discovered that the AI models incorrectly answered more than 60 percent of queries about news sources….The study highlighted a common trend among these AI models: rather than declining to respond when they lacked reliable information, the models frequently provided confabulations…” Ars Technica’s Benj Edwards reports on a study from researchers at Columbia that found a substantial rate of error in citation across multiple large language model AIs that are seeing routine use for search and research on the web.
- “TextGrad is built on three principles: (1) it is a general and performant framework that is not handcrafted for a specific application domain; (2) it is easy to use, mirroring PyTorch abstractions, thus allowing knowledge transfer; and (3) it is fully open source. Through TextGrad, we obtained state-of-the-art results in code optimization and PhD-level question answering, optimized prompts, and provided proof-of-concept results in optimizing molecules and treatment plans.” A research article published in Nature by Yuksekgonul and colleagues introduces TextGrad, a framework that employs improves large-language model output through the error-correction technique known as back propagation.
BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH
- “The little-known notebooks of the Victorian scientist Michael Faraday have been unearthed from the archive of the Royal Institution and are to be digitised and made permanently accessible online for the first time…The notebooks include Faraday’s handwritten notes on a series of lectures given by the electrochemical pioneer Sir Humphry Davy at the Royal Institution in 1812. ‘None of these notebooks have been looked at or analysed in any great depth,’ said Charlotte New, head of heritage for the Royal Institution. ‘They’re little known to the public.’” The Guardian’s Donna Ferguson reports that a trove of personal notebooks used by physics giant Michael Faraday have been digitized and will be made publicly available.
- “Online addictive behaviors are not an individual choice but may be a consequence of difficulties in making advantageous choices and regulating rewarding behaviors. To optimize prevention and treatment of these emerging mental health conditions, more global collaboration networks among research institutes are needed to realize large-scale multi-methodological longitudinal studies.” A perspective article published in Science by Matthias Brand examines the question of whether certain modes of internet use constitute an actual addiction.
- “To resolve inconsistencies between the genetic paradigm of cancer and biological reality, we must complement deep sequencing with deep thinking: embrace formal theory and historicity of biological entities, and (re)consider non-genetic plasticity of cells and tissues. In this Essay, we discuss the concepts of cell state dynamics and tissue fields that emerge from the collective action of genes and of cells in their morphogenetic context, respectively, and how they help explain inconsistencies in the data in the context of SMT [somatic mutation theory].” A perspective article published in PLOS Biology by Huang and colleagues argues for a shift in the ways cancer is currently conceived of as a genetic disease arising from somatic mutation (H/T @erictopol.bsky.social).
- “Among community-dwelling persons with moderate to severe dementia and their caregivers, a dementia care management program with integrated palliative care, compared with usual care, did not significantly improve patients’ neuropsychiatric symptoms or secondary outcomes of patient or caregiver symptoms through 24 months. However, it reduced the prespecified secondary outcome of combined ED and hospitalization events.” In an article published in JAMA, Sachs and colleagues present findings from a randomized trial of palliative vs usual care for participants with moderate to severe dementia and their caregivers.
- “…our results indicate low H5N1 virus infection risk through wastewater but higher potential risk from exposure to contaminated milk. Research into transmission risk through different pathways and the probability of infection arising from different doses and routes of exposure to H5N1 is ongoing, but our results call for particular caution in handling and disposing of milk from infected cattle.” In a dispatch article published in Emerging Infectious Diseases, Kaiser and colleagues present findings showing that highly pathogenic avian influenza (H5N1) can be detected in irradiated raw milk, raising the possibility that exposure to contaminated milk might present transmission risks.
COMMUNICATION, Health Equity & Policy
- “From ‘Envy, inequality and fertility’ to ‘Market size, trade and productivity’, using catchy three-part phrases in the titles of research papers can boost their citations, suggests a study…The study used algorithms to examine more than 235,000 economics papers and 93,000 medical and life-sciences papers that contain three-part phrases in their titles. The medical and life-sciences studies that used the format attracted 32 extra citations, on average, than did papers that don’t contain such phrases, and the economics papers with this format received an extra 3.5 citations.” Rule of 3: Nature’s Dalmeet Singh Chawla reports on a preprint available at SSRN that found using “tripartite” phrases in journal article titles was correlated with citation advantages for those papers.
- “…Oxford University Press (OUP) product strategy director John Campbell explained how the launch of Google’s AI-powered search summaries had led to a 19 per cent drop in click-through to academic reference services…Discussing what he called the advent of “zero-click journeys” for academic publishers, Campbell explained how half of all Google keyword searches likely to surface information within OUP’s own platform, Oxford Academic, now appeared with an AI-generated description next to them.” At Times Higher Education, Jack Grove reports on recent drops in visits to academic publishers’ websites as AI-created summaries usurp traffic.
- “Vancouver and many other districts around the country have turned to technology to monitor school-issued devices 24/7 for any signs of danger as they grapple with a student mental health crisis and the threat of shootings…The goal is to keep children safe, but these tools raise serious questions about privacy and security — as proven when Seattle Times and Associated Press reporters inadvertently received access to almost 3,500 sensitive, unredacted student documents through a records request about the district’s surveillance technology.” An article co-reported by the Seattle Times’ Claire Bryan and AP’s Sharon Lurye examines some potentially fraught complications of deploying pervasive AI monitoring of students on school technology platforms.
- “…it turns out how people engage with health tech depends a lot on their age cohort. From Gen Z to the Silent Generation, generations of Americans are using digital health in their own unique ways. For companies trying to engage across generations, differentiating between their habits and preferences is key.” A Rock Health infographic based on 2024 survey data provides a snapshot of how different generations are using (or not using) digital technology for accessing, managing, and interacting with healthcare (H/T @smcgrathphd).
