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.

January 31, 2025

In this week’s Duke AI Health Friday Roundup: DeepSeek model release makes big splash in the AI pond, turns heads; lifestyle interventions to prevent cognitive decline; time to ditch USMLE as health AI benchmark?; relationships between education and life expectancy across US counties; junk science pumped out by paper mills permeates search results; figuring out LLMs’ place in the scientific enterprise; setting priorities for health AI; more:

AI, STATISTICS & DATA SCIENCE

The flukes of a whale, visible as it dives beneath smooth water in a sound. Image credit: Thomas Lipke/Unsplash
Image credit: Thomas Lipke/Unsplash
  • “DeepSeek’s success is even more remarkable given the constraints facing Chinese AI companies in the form of increasing US export controls on cutting-edge chips. But early evidence shows that these measures are not working as intended. Rather than weakening China’s AI capabilities, the sanctions appear to be driving startups like DeepSeek to innovate in ways that prioritize efficiency, resource-pooling, and collaboration.” The AI world is still reverberating from the news about the reportedly impressive performance of the Chinese DeepSeek AI, which offers competitive performance with other leading-edge models at a fraction of the cost or need for compute resources. MIT Technology Review provides an overview of recent developments, while Gary Marcus provides perspective on some of the potential implications. Ars Technica reports on the apparent exposure of sensitive data in DeepSeek, and the original preprint paper describing DeepSeek is also available from arXiv.
  • “In this paper, we studied the impact of class imbalance corrections on the out-of-sample predictive performance of clinical prediction models developed with commonly used machine learning algorithms. We found that when data exhibited class imbalance, implementing imbalance corrections often led to deteriorated model calibration and (consequently) deteriorated overall performance. For both moderate (event fraction = 0.2) and strong imbalance scenarios (event fraction = 0.02), we found that correcting for class imbalance with a data pre-processing technique (RUS, ROS, SMOTE, SENN) and/or an imbalance-correcting algorithm (RB, EE) resulted in prediction models that consistently over-estimated risk.” In an article published in Statistics in Medicine, Carriero and colleagues evaluate whether machine-learning risk prediction models should be corrected for class imbalances.
  • “Medical exams are appealing as benchmarks due to their availability and the ease of evaluating multiple-choice questions. However, we argue that using these exams as benchmarks to validate LLMs for clinical tasks is not only misguided but a distraction from achieving the progress required for their safe, effective use in health care.” An editorial published in NEJM AI by Raji and colleagues suggests that the time for using medical licensing exams (such as USMLE) as a benchmark for LLM performance has come and gone.
  • “While the building blocks alone are not sufficient to solve these challenges, they are a necessary prerequisite. A health care system in which high-quality EHR data serve as the criterion standard source of truth cannot be realized if data remain trapped in individual EHR systems. Seamless health data exchange is of fundamental importance to ensure that health data can power both care and secondary data uses, such as business operations, research (with appropriate consent), and public health reporting, without imposing additional burdens on the health care workforce.” A JAMA Special Communication by Abassi and colleagues takes stock of the Department of Health and Human Services’ efforts to create a unified system for health data exchange.

BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH

Photograph of three runners in silhouette against a blue sky and bright sun. Image credit: Fitsum Admasu/Unsplash
Image credit: Fitsum Admasu/Unsplash
  • “Our randomized controlled trial of 6,104 dementia-free older adults was effective at improving global cognition over 3 years using a personalized, online, multidomain platform targeting modifiable dementia risk factors. The intervention benefited global cognition as well as component cognitive domains in at-risk older adults. Notably, positive health improvements across a range of lifestyle and dementia risk factors indicate the potential to help prevent cognitive decline over time.” A research article published in Nature Medicine by Brodaty and colleagues reports findings from a randomized trial examining lifestyle interventions designed to prevent cognitive decline.
  • “We show that prior vaccination can profoundly alter activation of innate immune responses. Although monocytes and NK cells from both infected groups were more activated compared with those of HCs, our findings revealed that monocytes and NK cells in breakthrough infection were less activated compared with their counterparts in primary infection. We further found that multiple communication pathways predicted to promote monocyte migration and NK cell proliferation were down-regulated during breakthrough infection compared with primary infection. The ability to clear infection without driving marked innate immune activation may be key to preventing severe disease, given that dysregulation of innate immune responses is strongly associated with COVID-19 disease severity.” A research article published in Science Translational Medicine by Chan and colleagues examines the effects of COVID vaccination on the innate immune response during breakthrough infections (H/T @EricTopol.bsky.social).
  • “Our results show large disparities across educational attainment populations, which persisted—and in many cases, grew—over time, and which were widespread geographically. The gap between the most and the least educated was substantial, reaching nearly 11 years in 2019. Concerningly, improvements in life expectancy during the first two decades of the 21st century were concentrated among the most highly educated, whereas those with less than a high-school diploma experienced no gains in life expectancy at all, exacerbating disparities. Geographical disparities were even larger, both within and across levels of education, and life expectancy for less educated Americans in some locations was remarkably low.” A research article published in Lancet Public Health by the Global Burden of Disease US Health Disparities Collaborators examines relationships between life expectancy and education in the US at the county level.

COMMUNICATION, Health Equity & Policy

Faceted rainbow lenses in a kaleidoscopic pair of novelty glasses. Image credit: Malcom Lightbody/Unsplash
Image credit: Malcom Lightbody/Unsplash
  • “How should the advancement of LLMs affect the practice of science? Do LLMs actually improve our scientific output or are they rather hindering good scientific practice? To what extent should they be used given the ethical and legal issues that come with them? We believe these to be highly non-trivial questions without an obvious answer and have therefore invited four groups of researchers to provide their perspectives on them.” A perspective article published in PNAS by Binz and colleagues provides multiple viewpoints on the appropriate use of large language models in scientific endeavors.
  • “Because Google Scholar is not an academic database, it is easy for the public to use when searching for scientific literature. That’s good. Unfortunately, it is harder for members of the public to separate the wheat from the chaff when it comes to reputable journals; even the difference between a piece of peer-reviewed research and a working paper can be confusing.” Gizmodo’s Isaac Schultz reports on recent research pointing to the growing prevalence of LLM-produced “junk science” in results pulled from Google Scholar searches.
  • “AI is poised to transform how patients, caregivers, and health care professionals experience the management of their health, health care, and care goals. Substantial challenges remain in realizing this promise. We believe that policies in four key areas can facilitate and accelerate AI in health and health care, including promoting the safe, effective, and trustworthy use of AI; promoting the development of an AI-competent health care workforce; focusing investments in key research portfolios; and clarifying AI liability and responsibilities.” An article published in Health Affairs by Matheny and colleagues presents perspectives from a National Academy of Medicine initiative on priorities for health AI.
  • “Paper mills flourish because of research systems that evaluate scientists using publication metrics, thereby inadvertently providing an incentive for misconduct. People with paper-mill publications might be promoted over those who have more modest — but honest — publication records….Institutions seldom seem to punish researchers for using paper mills, perhaps owing to a lack of awareness, or concern about reputational or legal risks.” A commentary article in Nature by Abalkina and colleagues proposes some steps for curbing the deleterious influence of “paper mills” that churn out low- or zero-quality scholarly publications.