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.

December 20, 2024

In this week’s Duke AI Health Friday Roundup: teaching AI ethics in grade school; checking on where all that training data is coming from; House task force releases AI report; NASEM analysis adds to debate on alcohol; examining the gap between lifespan and ‘healthspan’; COMET trial shines light on treatment for DCIS; proposing ethical guidelines for posthumous authorship; flood of low-quality commentaries engulfing some scholarly journals; much more:

AI, STATISTICS & DATA SCIENCE

“Educators may be willing to embrace the concept of ethical AI classes if given the resources, but there’s no clear framework on how that could be done. Should they begin teaching it in elementary school? Will it replace a class, be offered as an elective or be integrated into the existing curriculum? There’s no one-size-fits-all solution…Many teachers have adopted ‘not in my classroom’ policies in response to these concerns. As a result, students likely won’t learn how to use these tools the right way. Many will take whatever inaccurate, unregulated knowledge they cobble together from peers, parents and the internet as the truth — which may have far-reaching impacts.” An essay by Zac Amos at Hacker Noon asks why, amid the contradictory pressures of schools being urged to adopt AI technologies and desperate attempts to head off AI-enabled cheating and other misbehavior, there has not been more of an emphasis on teaching ethical aspects of using AI technologies?
“…transformers, the architecture underpinning language models, were invented in 2017, and the AI sector started seeing performance get better the bigger the models and data sets were. Today, most AI data sets are built by indiscriminately hoovering material from the internet. Since 2018, the web has been the dominant source for data sets used in all media, such as audio, images, and video, and a gap between scraped data and more curated data sets has emerged and widened…The past few years have also seen the rise of multimodal generative AI models, which can generate videos and images. Like large language models, they need as much data as possible, and the best source for that has become YouTube….For video models…over 70% of data for both speech and image data sets comes from one source.” An article by MIT Technology Review’s Melissa Heikkilä and Stephanie Arnett scrutinizes the evolving needs for data to train ever-larger transformer models has led to a few large (and increasingly powerful) data sources dominating the landscape.
“…we use state-of-the-art LLMs to generate three distinct complete versions of academic papers for each signal. The different versions include creative names for the signals, contain custom introductions providing different theoretical justifications for the observed predictability patterns, and incorporate citations to existing (and, on occasion, imagined) literature supporting their respective claims. This experiment illustrates AI’s potential for enhancing financial research efficiency, but also serves as a cautionary tale, illustrating how it can be abused to industrialize HARKing (Hypothesizing After Results are Known).” A not-yet-peer-reviewed preprint by Novy-Marx and Velikov, available from SSRN, demonstrate a method for generating surface-plausible research papers via LLM at large scale.
“The continual evolution in AI capabilities and integration has raised new policy issues. Some of the most prominent challenges involve data availability and quality, incomplete or inaccurate responses, non-individualized recommendations, decision transparency, data privacy and cybersecurity, interoperability between existing systems and AI, liability for errors made or enabled by AI models, and biased decision-making as well as the deployment of these models in a way that promotes financial gain over patient care and safety.” The Bipartisan House Task Force on AI has released a report containing guidance, principles, and policy recommendations for the use of AI technologies across a wide array of applications, including healthcare.

BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH

“For all-cause mortality, the panel’s review found strong evidence that heavy drinking has adverse effects on health, including increasing the risk of the leading causes of death. ‘However, the association of moderate alcohol consumption with all-cause mortality is less clear,’ the report says, citing evidence from eight eligible studies. According to its meta-analysis, the committee found those who consumed moderate levels of alcohol had a 16% lower risk of all-cause mortality than those who never drank. This conclusion was graded as being of moderate certainty, meaning there was enough evidence to determine the health effects, but there are limitations due to the quality of the evidence.” STAT News’ Isabella Cueto reports on the release of a long-awaited National Academies analysis that may further roil the ongoing scientific debate about the health effects of alcohol consumption.
“This study reported life expectancy and health-adjusted life expectancy trends over the past 2 decades and found healthspan-lifespan gaps for each of the 183 WHO member states. Sex disparities in healthspan-lifespan gaps and association with longevity and disease burden are also reported. The US stands out with the largest healthspan-lifespan gap and the greatest noncommunicable disease burden. The risk to healthspan is found amplified by longevity and is here recognized to be more pronounced in women.” An analysis by Garmany and Terzic published in JAMA Network Open illuminates the degree of divergence between lifespan and “healthspan” across the globe.
“In an intention-to-treat analysis, we found that the 2-year cumulative rate of invasive cancer was 5.9% for women randomized to guideline-concordant care and 4.2% for active monitoring. These results show that at 2 years, patients randomized to active monitoring have noninferior invasive breast cancer risk in the affected breast compared with those randomized to guideline-concordant care. The findings are novel, as all current treatments for DCIS require surgical excision, despite a growing body of evidence that supports that not all DCIS is destined to progress to invasive cancer.” In an article published in JAMA by Hwang and colleagues, investigators from the COMET randomized trial report results from a study that compared outcomes from women with low-risk ductal carcinoma in situ (DCIS) treated with and without endocrine therapy.
“A species of lungfish found in South America has claimed the title of the animal with the biggest genome sequenced so far. The DNA of Lepidosiren paradoxa comprises a staggering 91 billion chemical letters or ‘bases,’ 30 times as many as the human genome, researchers report today in However, those 91 billion bases of DNA only contain about the same number of genes that humans have—roughly 20,000—with the rest consisting of noncoding, perhaps even ‘junk’ DNA.” Science’s Elizabeth Pennisi reports on the recent discovery that a species of lungfish has an unexpectedly, inordinately large genome – and currently sits atop the leaderboard for largest animal genomes.

COMMUNICATION, Health Equity & Policy

“…despite interest in posthumous literary publication, little has been written about it in scholarly science. Calls for a clear policy have not resulted in uniform comprehensive guidelines, and occasional online discussions have shown that there is no consensus about how deceased colleagues should be credited in publications: some advocate coauthorship; others acknowledgment only. Nor are there clear guidelines on how an author’s death should be noted.” In an article published in BMJ, Nunan and Aronson address ethical issues raised by a lack of consensus about how to deal with posthumous publication in scientific research.
“…we show that (i) misinformation sources evoke more outrage than do trustworthy sources; (ii) outrage facilitates the sharing of misinformation at least as strongly as sharing of trustworthy news; and (iii) users are more willing to share outrage-evoking misinformation without reading it first.” A research article published in Science by McLoughlin and colleagues present findings from a study that reveals the role played by reactions of outrage and disgust in driving the sharing of misinformation online.
“Neurosurgical Review is not the only journal overwhelmed by commentary articles, a joint investigation by Science and Retraction Watch finds. They now make up 70% of the content in Elsevier’s Oral Oncology Reports, and nearly half in Wolters Kluwer’s International Journal of Surgery Open. At Neurosurgical Review, a Springer Nature title, letters, comments, and editorials comprised 58% of the total output from January to October—up from 9% last year…Science and Retraction Watch’s investigation suggests authors, journals, and institutions all benefit from the scheme, which floods the literature with poor-quality publications and casts doubt on metrics of scholarly output and impact.” Science Insider reports on a joint Science-Retraction Watch investigation that reveals an onslaught of shoddy commentary publications in scholarly journals.