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.
July 21, 2023
In this week’s Duke AI Health Friday Roundup: a primer on foundation models; genetics and asymptomatic COVID; overblown claims for AI content detection; don’t trust GPT with the baby just yet; AI thirst for data drives interest in synthetic sources; the merits of working (out) for the weekend; physics offers window on sudden heart arrhythmias; tracing developments in press coverage of scientific preprints; expanding vaccination coverage for uninsured adults; much more:
AI, STATISTICS & DATA SCIENCE
- “Clinicians, educators, and trainees are eager to integrate these tools into clinical workflows to improve the delivery of care. However, our results suggest that LLMs currently lack the ability to reliably generate neonatal-perinatal board examination questions and do not consistently provide explanations that are aligned with the scientific and medical consensus….it is unlikely that it would pass the actual neonatal-perinatal medicine board examination and should not be considered a reliable resource for maintenance of certification examination.” A research letter published in JAMA Pediatrics by Beam and colleagues examines how well the GPT-3.5 transformer large language model performs on board exam questions for neonatal pediatrics.
- “An emerging type of AI system is a ‘foundation model’, sometimes called a ‘general-purpose AI’ or ‘GPAI’ system. These are capable of a range of general tasks (such as text synthesis, image manipulation and audio generation)….Because foundation models can be built ‘on top of’ to develop different applications for many purposes, this makes them difficult – but important – to regulate. When foundation models act as a base for a range of applications, any errors or issues at the foundation-model level may impact any applications built on top of (or ‘fine-tuned’) from that foundation model.” New from the Ada Lovelace foundation: a handy, nontechnical explainer that covers the basics of what “foundation” model AIs (such as GPT-*) are, how they work, and what all that arcane terminology actually means.
- “…our investigation underscores the power of incorporating human and social factors to produce AI that complements rather than substitutes for human expertise…When AI hypothesis generation is made aware of human expertise, it can accelerate discovery and liberate human scientists to steer science and technology in novel directions. Our system and its recommendations raise ethical concerns; they could be used as a ‘scoop-machine’ to leapfrog human scientists and seize on answers that they might otherwise ask and answer next.” An article published in Nature Human Behavior by Sourati and Evans proposes an approach for tuning AIs to predict future discoveries via “hypergraph proximity.”
- “But as generative AI software becomes more sophisticated, even deep-pocketed AI companies are running out of easily accessible and high-quality data to train on. Meanwhile, they are under fire from regulators, artists and media organisations around the world over the volume and provenance of personal data consumed by the technology.” An article by Financial Times’ Madhumita Murgia examines the AI industry’s growing interest in synthetic data for training models.
BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH
- “Specifically, these findings suggest that engagement in physical activity, regardless of pattern, may optimize risk across a broad spectrum of cardiovascular diseases….efforts to increase physical activity for cardiovascular health may be effective even when such efforts are concentrated into 1 to 2 days per week. Because weekend warrior patterns may be more feasible for certain schedules, targeted interventions delivered over shorter timeframes may be more accessible.” A research article published this week in JAMA by Kurshid and colleagues describes results from a study that assessed the effects of episodic weekend exercise (vs more evenly distributed exercise) on mortality and other cardiovascular outcomes.
- “He’s hoping to find a way to stop arrhythmias without having to use traditional defibrillator paddles that send a huge blast of electricity through the patient’s entire body. Instead, Fenton is trying to fight waves with waves. He’s making waves of his own to snuff out the pernicious spiral waves that can send a heart into disarray. The goal is to find a gentler, less damaging way to treat arrhythmias.” An episode of Quanta’s Joy of Why podcast delves into the mechanisms behind sudden heart arrhythmias from a physics perspective as host Steven Strogatz talks with physics professor Flavio Fenton (text transcript available).
- “We show that, among participants reporting a positive test result for SARS-CoV-2, HLA-B*15:01 is significantly associated with asymptomatic infection. We observed that individuals carrying this common allele (approximately 10% in individuals with European ancestry) are more than twice as likely to remain asymptomatic after SARS-CoV-2 infection compared with those who do not, and a notable effect of HLA-B*15:01 homozygosity increasing the chance of remaining asymptomatic by more than eight times.” A research paper published by Augusto and colleagues in this week’s Nature sheds light on the role of genetics in determining whether a person who contracts COVID develops symptoms.
- “In a trial of the MIND diet that was designed to improve brain health, cognitive function and brain imaging outcomes at 3 years did not differ significantly between participants who followed the MIND diet and those who followed a control diet with a mild caloric restriction.” A paper published in the New England Journal of Medicine by Barnes and colleagues presents results from a randomized trial that assessed the protective effects of a diet that combines elements of the Mediterranean and DASH diets on the development of Alzheimer disease in patients without evidence of cognitive decline but with a family history of dementia.
COMMUNICATION, Health Equity & Policy
- “…fragmentation, for human readers, is mostly a massive inconvenience. For the machines, though, whether for training or analysis, fragmentation is a real impediment. The machines never, ever, want to read the publications just of a single publisher. The machines want to read everything, or at least everything in a given field. And this in turn raises questions about the work of assembling everything together — which is no small task — and making it available for machine training and analysis.” A post at Scholarly Kitchen by Roger C. Schonfeld offers a detailed look at how the advent of AI in scholarly spaces is affecting the ways in which online content is assembled and made available.
- “…the rise in preprint coverage was most pronounced among health and medicine-focused media outlets, which barely covered preprints before the pandemic but mentioned more COVID-19 preprints than outlets focused on any other topic. These results suggest that the growth in coverage of preprints seen during the pandemic period may imply a shift in journalistic norms, including a changing outlook on reporting preliminary, unvetted research.” A research article by Fleerackers and colleagues, available as a preprint from BioRxiv, examines patterns in recent media coverage of scientific article preprints.
- “With the proliferation of widely available generative AI tools has come a commensurate rise in detection tools marketed as capable of identifying generated content. Some of these tools may work better than others. Some are free and some charge you for the service. And some of the attendant marketing claims are stronger than others – in some cases perhaps too strong for the science behind them.” A post at the Federal Trade Commission’s blog by FTC attorney Michael Atleson turns a skeptical eye on products being marketed for the detection of AI-generated content (H/T @GaryMarcus).
- “Vaccine costs are a well-recognized barrier to access, are often highest for uninsured people, and contribute to disparities in adult vaccination rates by insurance status. But the national free-to-patients Covid-19 vaccination program demonstrated that out-of-pocket costs aren’t the only impediment to vaccination; limited numbers of vaccination sites in rural settings and areas with low concentrations of primary care providers and failure to disseminate straightforward, trusted information about vaccines also hinder uptake among adults.” A perspective article published in the New England Journal of Medicine by Wallender and colleagues argues for expanded efforts to provide vaccination coverage for uninsured adults.
