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 3, 2025
In this week’s Duke AI Health Friday Roundup: pumping the brakes on “mirror life” experiments; talking federated registration for health AI; state of play for H5N1 infections in humans; privacy challenges for synthetic data; new access rules for federally funded research to go into effect; learning from longitudinal digital health; who owns the rights to your digital twin; can we build better LLMs with retrieval-augmented generation?; more:
AI, STATISTICS & DATA SCIENCE
- “The reality is that the pace of the AI revolution has been much faster than our neatly designed process. It’s too much for our organization to evaluate every AI technology that we think we might want to employ. So we said, ‘Well, okay, we need to do something different. We need to decide which tools are higher risk.’ We might apply more scrutiny to a tool that touches the patient, such as one that predicts the risk of a serious condition, than we would something that is maybe creating a note for a clinician to review. We cannot decide how much scrutiny to apply if we don’t know that we have it.” Duke AI Health Director Michael Pencina is interviewed by JAMA on the prospects for developing a federated registration system for health AI.
- “Designers of prospective digital health studies need to understand how participants engage to improve data collection methods. Modelling methods that are appropriate for longitudinal data with varying amounts of missingness can guide the development of strategies to improve study engagement and garner representative datasets. The evidence suggests that demographic factors, including age, socioeconomic status, and digital literacy, can affect retention and adherence rates.” A research article by Cho and colleagues, published in Lancet Digital Health, sifts through data from a pair of longitudinal digital health studies to extract lessons for future study design.
- “Synthetic data are not inherently free from reidentification risks. When based on real patient data, synthetic data may overfit (replicating actual records) or be vulnerable to reidentification through techniques such as membership inference (detecting whether a model used someone’s data by analyzing outputs) or linkage attacks (matching quasi-identifiers across datasets). Currently, no widely accepted standards exist for generating or evaluating synthetic health care data, and existing metrics alone remain insufficient. Evaluations should comprehensively assess fidelity, utility, and privacy, addressing inherent trade-offs among them.” A Viewpoint article by Abrall and colleagues, appearing in this week’s JAMA, explores potential health privacy issues that may emerge from the use of synthetic data for training health AI applications.
- “LLMs, as a subset of generative AI, have demonstrated value in content creation, idea generation, and interactive communication. However, their inherent limitations, such as the need for up-to-date information, hallucinations of incorrect facts, and a reliance on public-domain data, restrict the full potential of generative AI within the health care setting. To address these limitations, retrieval-augmented generation (RAG) offers a novel framework by connecting LLMs with external knowledge, enabling them to access information beyond their training data.” A review article in NEJM AI by Karen Ka Yan Ng and colleagues examines the use of retrieval-augmented generation for improving the reliability and trustworthiness of large language models in healthcare applications.
BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH
- “All of life’s primary biomolecules can exist in two mirror-image forms, like a left and right hand. But only one form is found in nature. Proteins are left-handed, for example, and DNA and RNA are right-handed. Synthetic biologists have previously synthesized mirror-image proteins and genetic molecules. And mirror-image amino acids and peptides—the building blocks of proteins—have been incorporated into several approved drugs….The concern…is that taking this line of work many steps further could result in fully mirror-image bacteria that could reproduce. Such organisms would likely be able to infect and potentially harm a wide range of microbes, plants, and animals while resisting the molecules that enable predators to kill and digest existing microbes.” At Science, Robert F. Service reports on recent unease among biologists at the prospect of creating synthetic life from biomolecules with reversed chirality.
- “The development of an effective influenza vaccine remains challenging. A major issue with the current vaccines is that most individuals respond better to the strain that they are biased for and thus may have little protection against infection by other strains. Here, we have shown that coupling heterologous HA antigens can limit this subtype bias in both mice and human tonsil organoids, most likely by increasing T cell help. We have also shown that T cell help can be ‘borrowed’ from the favored seasonal strain to boost antibody response to an avian strain in tonsil organoids.” A research article by Mallajosyulla and colleagues, published in Science, suggests new avenues for improving the effectiveness of influenza vaccines.
- “In the cases identified to date, A(H5N1) viruses generally caused mild illness, mostly conjunctivitis, of short duration, predominantly in U.S. adults exposed to infected animals; most patients received prompt antiviral treatment. No evidence of human-to-human A(H5N1) transmission was identified. PPE use among occupationally exposed persons was suboptimal, which suggests that additional strategies are needed to reduce exposure risk.” A research article published in the New England Journal of Medicine by Garg and colleagues provides a snapshot of outcomes from humans infected with the highly pathogenic strain of avian H5N1 influenza.
COMMUNICATION, Health Equity & Policy
- “While some fear that GenAI tools will replace people so that no one will have jobs any longer, I have a slightly more subtle concern. What makes someone valuable to a team is the skills they bring to the table. If this knowledge can be extracted, stored, and regenerated in novel ways in new situations, does one ever “leave” at the end of one’s employment? What rights does your former employer have to continue to “use” your agent after you quit or retire?” An essay by Todd A. Carpenter at Scholarly Kitchen muses on the potential legal and ethical complexities of AI agents trained on actual human personalities – particularly when that takes place in the workplace.
- “The full proposed rule was posted to the Federal Register on Friday, and Department of Health and Human Services posted a more condensed breakdown on its website….The healthcare information of more than 167 million people was affected in 2023 as a result of cybersecurity incidents, she said….The proposed rule from the Office for Civil Rights (OCR) within HHS would update standards under the Health Insurance Portability and Accountability Act (HIPAA) and would cost an estimated $9 billion in the first year, and $6 billion in years two through five…” Reuters’ A.J. Vincens reports on a new proposed federal rule that would tighten cybersecurity requirements in the healthcare industry.
- “Two years after President Joe Biden’s administration shook up scientific publishing by calling for immediate free access to scientific journal articles produced from federally funded research by the end of 2025, the U.S. National Institutes of Health (NIH) and Department of Energy (DOE) have released their final plans for complying. All other U.S. research funding agencies are expected to follow suit by the end of this month….The NIH and DOE policies require grantees to post accepted, peer-reviewed manuscripts in each agency’s public repository as soon as they are published, among other stipulations. Research funding agencies are also expected to require immediate sharing of project data.” Science reports on the looming enactment of government policies that will affect access to published science and underlying data from federally funded research.
