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 13, 2025

In this week’s Duke AI Health Friday Roundup: “juggling” increased work demands with AI may incur steep human costs; chatbot responses to health questions are highly sensitive to phrasing; statin study meta-analysis finds no causal link with commonly reported adverse events; analysis of educational AI finds quality of evidence for outcomes lacking; “digital twin” approach to limb prosthetic control; much more:

AI, STATISTICS & DATA SCIENCE

A man wearing a white suit is juggling balls. Three are in the air above his head; two are in his hands. Image credit: Marco Bianchetti/Unsplash
Image credit: Marco Bianchetti/Unsplash
  • “Once the excitement of experimenting fades, workers can find that their workload has quietly grown and feel stretched from juggling everything that’s suddenly on their plate. That workload creep can in turn lead to cognitive fatigue, burnout, and weakened decision-making. The productivity surge enjoyed at the beginning can give way to lower quality work, turnover, and other problems.” A report at Harvard Business Review by researchers Aruna Ranganathan and Xingqi Maggie Ye examines the evidence for purported AI productivity boosts and finds some problematic aspects to the phenomenon.
  • “Insight into the incidents comes as AI is beginning to transform the world of health care. Proponents predict the new technology will help find cures for rare diseases, discover new drugs, enhance surgeons’ skill and empower patients. But a Reuters review of safety and legal records, as well as interviews with doctors, nurses, scientists and regulators, documents some of the hazards of AI in medicine as device makers, tech giants and software developers race to roll it out.” Reuters’ Jaimi Dowdell, Steve Stecklow, Chad Terhune and Rachael Levy examine an increase in reports of adverse events associated with the use of certain AI-enabled medical devices.
  • “The experiment found that the chatbots were no better than Google — already a flawed source of health information — at guiding users toward the correct diagnoses or helping them determine what they should do next. And the technology posed unique risks, sometimes presenting false information or dramatically changing its advice depending on slight changes in the wording of the questions.” The New York Times’ Teddy Rosenbluth reports on research recently published in Nature Medicine by Bean and colleagues, in which the authors found that chatbot performance on answering medical questions is sensitive to how the user frames the question, suggesting that testing with real-world scenarios may be needed.
  • “We found that using AI assistance led to a statistically significant decrease in mastery. On a quiz that covered concepts they’d used just a few minutes before, participants in the AI group scored 17% lower than those who coded by hand, or the equivalent of nearly two letter grades. Using AI sped up the task slightly, but this didn’t reach the threshold of statistical significance…Importantly, using AI assistance didn’t guarantee a lower score. Howsomeone used AI influenced how much information they retained.” An observational study performed by AI developer Anthropic found that the increased productivity reported with AI-assisted computer coding may come with a serious trade-off: reliance on AI assistance led to a significant decrease in skills mastery.

BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH

Along with blood vessels (red) and nerve cells (green), this mouse brain shows abnormal protein clumps known as plaques (blue). These plaques multiply in the brains of people with Alzheimer's disease and are associated with the memory impairment characteristic of the disease. Because mice have genomes nearly identical to our own, they are used to study both the genetic and environmental factors that trigger Alzheimer's disease. Experimental treatments are also tested in mice to identify the best potential therapies for human patients. This image was part of the Life: Magnified exhibit that ran from June 3, 2014, to January 21, 2015, at Dulles International Airport. Image credit: Alvin Gogineni, Genentech via NIGMS
Image credit: Alvin Gogineni, Genentech via NIGMS
  • “…we established the feasibility of amyloid plaque–specific CAR-T cells as a potential therapeutic avenue for AD [Alzheimer disease]. These findings highlight the potential of CD4+ CAR-T therapy not only to modify amyloid pathology but also to reshape the immune landscape of the CNS, paving the way for future development of cellular immunotherapies for neurodegenerative disease.” A research article published in PNAS by Boskovic and colleagues offers a preliminary demonstration for the application of CAR-T immunotherapy to modify the deposition of amyloid deposits in a mouse model of Alzheimer disease.
  • “After brief training, participants learned to modulate muscle activation timing and amplitude to control prosthetic joint torques with no need for state machines or rule-based controllers. These results show that our human-machine interface, based on an electromyography-driven musculoskeletal model, can generalize myoelectric control across movements and amputation types, representing a step toward more natural, adaptive, and user-driven bionic limbs.” A research article published in PNAS Nexus by Damonte and colleagues reports on the use of a “digital twin” to develop better control for prosthetic limbs by using a patient’s intact limb to model the bionic limb’s responses.
  • “Ekstrøm and colleagues examined whether results for other BHP trials were missing. …they identified dozens of trials carried out at sites around the project’s base in Guinea-Bissau and whittled the sample down to any that were labeled as completed, terminated, or had an “unknown” status in the registry. Then they looked at whether the outcomes had been made public…For seven trials, including the DTP study now described in a preprint, Ekstrøm’s team could only find publications discussing some, rather than all, of the planned primary outcomes as described in the trial registry. Science’s Catherine Offord reports on recently published allegations that a number of clinical trials by an already-controversial group of Danish researchers responsible for a study of infant immunization against hepatitis B, conducted in Guinea-Bissau, may be missing data on study outcomes.
  • “Adverse event data from blinded randomised trials do not support causal relationships between statin therapy and most of the conditions (including cognitive impairment, depression, sleep disturbance, and peripheral neuropathy) listed in product labels as potential undesirable effects. In light of these findings, such labelling and other official sources of health information should be revised so that patients and their doctors can make appropriately informed decisions regarding statin therapy.” A meta-analysis published in Lancet by the Cholesterol Treatment Trialists’ Collaboration examines the evidence for causal links between the use of statin therapies and adverse events often associated with the widely used drugs.

COMMUNICATIONS & Policy

A young girl takes the extended hand of a humanoid robot. Image credit: Andy Kelly/Unsplash
Image credit: Andy Kelly/Unsplash
  • “The bot is so good at finding ways to talk to patients. Doctors also know it is so good at diagnosing and so good at reading scans and images — better than many doctors, in fact — and so good at answering patient questions in portals and writing appeals to insurance companies when a medication or procedure is denied…So what is a doctor for?” The New York Times’ Gina Kolata examines the current and likely future impact of AI on the medical profession.
  • “The results show strong evidence of severe publication bias and extreme between-study heterogeneity.…Subgroup analyses fail to identify specific moderators that yield consistent benefits. No substantial difference exists between studies published before or after 2023. Overall, broad claims of generalized learning gains resulting from AI/LLMs appear premature…” A preprint article by Bartos and colleagues, available from PsyarXiv, reports findings from a “meta-meta analysis” that show problematic patterns in the literature evaluating the effectiveness of AI-powered pedagogical tools.
  • “Given a question like ‘What are ways to cool the center-of-mass motion of levitated nanoparticles?’ OpenScholar responds by checking a database of 45 million open-access papers optimized for searches about science….Unlike earlier LLMs, which typically provided answers drawn from only one paper at a time, it examines content from multiple relevant papers. …OpenScholar’s answers run several hundred words longer than those produced by other models, helping it capture more nuance useful to scientists.” Science’s Jeffrey Brainard reports on findings, published in Nature by Asai and colleagues, from a study evaluating the performance of OpenScholar, an AI application designed specifically for the scientific and research communities.

Sites@Duke Express is powered by WordPress. Read the Sites@Duke Express policies and FAQs, or request help.