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.
February 7, 2025
In this Friday’s Duke AI Health Friday Roundup: MASAI study tests AI for mammography screening; the growing health threat posed by microplastics and an association with dementia; LLM “translates” clinic notes into plain-language text; global survey on AI use in science; the indirect toll of morbidity and mortality exacted by flooding disasters; adopting zero-trust architecture for scholarly publishing; best practices for deploying trustworthy health AI; much more:
AI, STATISTICS & DATA SCIENCE
- “The dirty little secret of Phase 1 in oncology is a Phase 1 study is not properly powered for anything. Certainly not for efficacy; borderline for safety, even though that’s the primary endpoint. A placebo-controlled randomized study is never run in Phase 1 in oncology, so the Phase 1 studies inevitably are single arm…. so we take guesses and we give it to statisticians and they throw their hands up and we muddle through, or having a reference cohort — so as if you had dosed the placebo group. But the reality is the placebo-dosed group in Phase 1 never happens. So it’s digital twins or it’s nothing.” STAT News’ Brittany Trang interviews one of the leaders of a French biotech company that is employing AI for drug development as part of an “end-to-end” application of the technology.
- “This paper describes the FUTURE-AI framework, which provides guidance for the development and deployment of trustworthy AI tools in healthcare. The FUTURE-AI Consortium was founded in 2021 and comprises 117 interdisciplinary experts from 50 countries representing all continents, including AI scientists, clinical researchers, biomedical ethicists, and social scientists….To operationalise trustworthy AI in healthcare, a set of 30 best practices were defined, addressing technical, clinical, socioethical, and legal dimensions. The recommendations cover the entire lifecycle of healthcare AI, from design, development, and validation to regulation, deployment, and monitoring.” A methods paper published in BMJ this week by Lekadir and colleagues debuts the FUTURE-AI consensus guidelines for deploying healthcare AI.
- “In this study, we utilized multimodal routine clinical data from 15,726 patients with solid cancers undergoing systemic treatment to uncover the complex mechanisms that determine a patient’s prognosis…By systematically comparing these AID [AI-derived] markers among patients, we show that prognostic associations are not static and that different markers may be critical depending on the cancer entity and the individual disease setting. In contrast to traditional statistical methods, xAI [explainable AI] can build on all available data to assess the complex setting of individual patients, provided that common pitfalls are addressed.” A research article published in Nature Cancer by Keyl and colleagues describes the training and evaluation of an explainable AI clinical decision-support system that analyzed clinical markers from real-world cancer patient data.
- “This analysis showed that a screen-reading procedure that used AI to triage screening examinations to single or double reading and that used AI as detection support in mammography screening led to a significant 29% increase in cancer detection compared with standard double reading without AI (6·4 vs 5·0 per 1000 participants screened), with a similar false-positive rate and with a substantial 44% reduction in the screen-reading workload. The increase in detection mostly concerned small, lymph-node negative, invasive cancers.” Lancet Digital Health has published a research article by Hernström and colleagues that reports findings from the MASAI trial, which evaluated the use of AI for screening mammography images for evidence of cancer.
- “This study demonstrates that GPT-4–based translation of DSNs resulted in large, statistically significant improvements in objectively measured comprehension, subjectively measured comprehension, and confidence. It also reduced the rate at which patients read DSNs. While paradoxical at first glance, this finding may suggest that they were more attentive or engaged when reading the translated content, leading to improved comprehension.” A study published in NEJM AI by Anivarya Kumar and a group of authors from Duke University evaluated the use of a GPT-4-based large language model that “translated” physicians’ clinical notes into plain text, with the goal of improving patient comprehension.
BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH
- “The essential point of this post is that the striking brain accumulation of MNPs is paralleled by the overwhelming accumulation of evidence for their toxicity to human health. It’s not just “innocent” MPN deposits in our brain and throughout our body—we have compelling data for MPN inciting an aggressive inflammatory response that has now been consistently documented across multiple organ systems. Yet we are not doing anything to take this on and it’s high time we get on it.” At his Ground Truths Substack page, Eric Topol unpacks findings from a recently published Nature study that reported some potentially alarming findings about the accumulation of micro- and nanoplastic particles in human bodies – including a possible association with dementia.
- “Using nearly two decades of highly resolved flood exposure data, we found that flooding events were associated with increased death rates for cardiovascular diseases, infectious and parasitic diseases, injuries and respiratory diseases, with generally greater associations for more severe floods compared with less severe. Injury death rates were higher for females and older adults associated with acute exposure to tropical cyclone-related flooding, and infectious disease mortality was elevated for males associated with exposure to all flood causes, particularly tropical cyclone- and heavy rain-related floods.” A research article published in Nature Medicine by Lynch and colleagues finds that large flooding events in the US are associated with changes in rates of mortality stemming from cancer, cardiovascular disease and other causes.
- “The main limitations of the evidence were the small numbers of people and lack of studies overall. We therefore have little confidence in the evidence because there were not enough studies to be certain about the results, and the included studies were very small; also, the evidence does not cover all the people we were interested in. Further research, including large studies, is needed to better understand how precision nutrition interventions may be used in the treatment or management of overweight and obesity in children and adolescents.” A Cochrane review by Huey and colleagues scrutinizes the evidence supporting precision-nutrition interventions for obesity in children and adolescents.
COMMUNICATION, Health Equity & Policy
- Time for a re-run: this Scholarly Kitchen article by David Crotty, originally published in 2018, is circulating again – and the points it makes about the need for careful and nuanced approaches to evaluating the “impact” of research are still salient: “Time scales come into play here as well. It takes about 12 years to go from the laboratory bench to an FDA approved drug. Most funders aren’t going to wait 12 years for you to prove societal impact to get your grant renewed. Remember also that only about 1 in every 5,000 drugs that enters preclinical testing will reach final approval. If your drug gets tested and fails, is that real world impact?…Exerting pressure to productize research, and to strive for small, short term gains over creating long term value is bad for research.”
- “…it is now technically possible to turn lead into gold, albeit far more expensive than it would be to just go buy gold directly. Unfortunately, and for different reasons, we are returning to an era when validation and verification need to become key aspects of our community. A variety of research integrity issues have been the focus of much attention. Scientific communication needs to return to an even more skeptical footing. We could do well to take a lesson from IT security and move scholarly communications to a zero-trust framework.” At Scholarly Kitchen, Todd Carpenter makes a case for adopting “zero-trust” principles from the world if IT security to ensure integrity in scholarly publishing.
- “Despite a burgeoning interest in AI tools, the survey suggests that researchers need more support to use them confidently. Nearly two-thirds of respondents said that a lack of guidance and training is preventing them from using AI to the extent that they would like….Researchers are also worried about how safe it is to use these tools: 81% of respondents said they had concerns about AI’s accuracy, potential biases, privacy risks and the lack of transparency in how these tools are trained.” Nature’s Miryam Naddaf reports on a recent global survey by scientific publisher Wiley that probed the scientific community about its use of AI in scientific research.
- “This case is hardly the first time we’ve seen a protracted retraction. Last year, for example, we covered the University of Maryland’s mission to retract a researcher’s faked data, and in 2020, philosophy professor Michael Dougherty detailed his “two year drama” to get a plagiarized paper retracted….In the six years it took for the papers to be retracted, the paper in FEBS Letters was cited 19 times and the BJ paper 14 times, according to Clarivate’s Web of Science. They have been cited 78 times and 71 times in total, respectively.” A post at Retraction Watch traces the tortuous and glacial process of getting a paper retracted for problems with some of the scientific images.
