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.
August 23, 2024
In this week’s Duke AI Health Friday Roundup: engineering vs cognitive science in AI; TACT2 reports results on chelation study in CVD, diabetes; a registry for health AI; trying to avoid past missteps with health tech evaluation; AI decision support and human cognition; database captures spectrum of risks from AI; safeguarding children’s digital determinants of health; scrutinizing cancer screening and health disparities; screening human microbiome for potential antimicrobials; much more:
AI, STATISTICS & DATA SCIENCE
- “…we set out to release the grip of the currently dominant view on AI (viz., AI-as-engineering aiming at a human-level AI system…). This practice has taken the theoretical possibility of explaining human cognition as a form of computation to imply the practical feasibility of realising human (-like or -level) cognition in factual computational systems; and, it is framing this realisation as a short-term inevitability….we undercut these views and claims by presenting a mathematical proof of inherent intractability (formally, NP-hardness) of the task that these AI engineers set themselves.” A preprint article by van Rooij and colleagues, available from PsyArXiv, argues that a purely engineering approach to developing human-level AI is fundamentally infeasible.
- “The team combed through peer-reviewed journal articles and preprint databases that detail AI risks. The most common risks centered around AI system safety and robustness (76%), unfair bias and discrimination (63%), and compromised privacy (61%). Less common risks tended to be more esoteric, such as the risk of creating AI with the ability to feel pain or to experience something akin to ‘death.’ …The database also shows that the majority of risks from AI are identified only after a model becomes accessible to the public. Just 10% of the risks studied were spotted before deployment. An article by Scott J. Mulligan at MIT Technology Review dives into a public database compiled by MIT CSAIL that captures the various ways AI could jump the tracks in ways large and small.
- “Presently, standards, regulations, and oversight lag the rapid advance of these technologies, presenting a 2-fold danger. On one hand, without meaningful oversight and evaluation, patients may be exposed to ineffective, biased, or even harmful technologies…On the other hand, adoption of potentially beneficial technologies may be hindered by a lack of trust that they will work as they should. More examples of successful deployments, including local context and granular operational details, are needed.” In a viewpoint article for JAMA, Duke AI Health director Michael Pencina and colleagues propose a registration system for health AI applications inspired in part by the ClinicalTrials.gov registry for clinical trials.
- “We argue that clinicians are contextually motivated, mentally resourceful decision makers, whereas AI models are contextually stripped, correlational decision makers, and discuss misconceptions about clinician–AI interaction stemming from this misalignment of capabilities. We outline how future research on clinician–AI interaction could better address the cognitive considerations of decision making and be used to enhance the safety and usability of AI models in high-risk medical decision-making contexts.” A viewpoint article published by Tikhomirov and colleagues in Lancet Digital Health offers a critical examination of how AI-based clinical decision support tools mesh – or don’t – with human cognitive processes.
- “In a field experiment involving nearly a thousand students, we have deployed and evaluated two GPT based tutors, one that mimics a standard ChatGPT interface (called GPT Base) and one with prompts designed to safeguard learning (called GPT Tutor). These tutors comprise about 15% of the curriculum in each of three grades. Consistent with prior work, our results show that access to GPT-4 significantly improves performance (48% improvement for GPT Base and 127% for GPT Tutor). However, we additionally find that when access is subsequently taken away, students actually perform worse than those who never had access (17% reduction for GPT Base). That is, access to GPT-4 can harm educational outcomes. These negative learning effects are largely mitigated by the safeguards included in GPT Tutor.” A preprint by Bastani and colleagues available at SSRN suggests that the integration of generative AI into education may backfire as students become dependent on it for certain tasks, unless care is taken to short-circuit that process.
BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH
- “…we argue that children’s use of digital technologies and engagement in digital environments should be recognised as important determinants of their health and that a public health approach is required to protect children from digital harms. Drawing on lessons from well-established approaches to address other public health challenges, we summarise three groups of public health interventions that can help delay media use among very young children, reduce digital media use among children of all ages, and mitigate any harmful consequences of children’s digital media use. In a viewpoint paper published in Lancet Public Health, Holly and colleagues make a case for early interventions aimed at “digital determinants of health” in young children.
- “The primary finding of this study is that the clinical benefits of EDTA chelation found in the earlier TACT trial were not replicated. Thus, the findings herein do not support the clinical use of chelation to reduce CVD risk in US and Canadian participants with diabetes and a prior MI.” In a research article published in JAMA, Lamas and colleagues present findings from the TACT2 randomized trial evaluating the use of chelation therapy for the prevention of major cardiovascular events in patients with diabetes and a history of myocardial infarction.
- “…we find sharp increases in sports betting following legalization. This increase does not displace other gambling activity or consumption but significantly reduces households’ savings allocations, as negative expected value risky bets crowd out positive expected value investments. These effects concentrate among financially constrained households, who become further constrained as credit card debt increases, available credit decreases, and overdraft frequency rises.” A research paper by Baker and colleagues, available as a preprint from SSRN, attempts to estimate the causal effects of online sports gambling on the health of household incomes.
- “…we curated a list of 323 high-confidence SEP families predicted to be expressed in the human microbiome and holding great promise as antimicrobials. We synthesized and tested 78 of these and found that more than half displayed antimicrobial activity against at least one pathogen or commensal. The active SEPs were subjected to detailed characterization to determine their mechanism of action, secondary structure, and toxicity toward human cell lines. Interestingly, the five most promising SEPs were encoded by diverse phyla from oral, skin, and gut body sites…” A research paper published in Cell by Torres and colleagues explores the use of computational screening of human microbiomes for candidate peptides that showed potential antimicrobial activity (H/T @EricTopol).
COMMUNICATION, Health Equity & Policy
- “While the desire to decrease health disparities is laudable, more cancer screening cannot accomplish that goal. Furthermore, the focus on cancer itself is too narrow, since less than 15% of the disparity in overall deaths between Black and white Americans is related to cancer. To genuinely address this disparity will take investing in poor communities regardless of their racial composition, though understanding that many are disproportionately Black. To move toward true health equity, interventions must prioritize community needs over system needs.” In an opinion article for STAT News, Adamson and colleagues question the emphasis placed on cancer screening as a means for rectifying racial health disparities in the US.
- “With the launch of a metascience unit late last year, the United Kingdom became one of the first countries to formalize the practice of using scientific methodology to study how research is done. The question now is whether it will produce insights to help elevate UK research and, if so, whether this will influence other countries to launch their own metascience initiatives…Its remit is to explore better ways of conducting, publishing and reviewing UK research, as well as distributing and funding it. More broadly, the unit is focused on improving the overall quality and efficiency of UK research.” Nature’s Dalmeet Singh Chawla reports on the United Kingdom’s launch of a government-funded initiative in the United Kingdom devoted to the topic of metascience – the study of how science gets done.
- “With increasing concern over the use of AI in clinical care, the combination of algorithm assurance testing and certification, along with a CMS requirement to require local testing and monitoring, will ensure that AI oversight is a shared responsibility of developers, users, and other stakeholders. Leveraging learnings from the health IT certification program can help avoid past pitfalls and better position an AI assurance testing and certification program for success.” A viewpoint article published in JAMA by Ratwani and colleagues proposes a set of recommendations for approaching the testing and certification of AI applications in healthcare – recommendations informed by salutary experiences with EHRs.
- “Even if it were possible to prove that an LLM has been trained on a certain text, it is not clear what happens next. Publishers maintain that, if developers use copyrighted text in training and have not sought a licence, that counts as infringement. But a counter legal argument says that LLMs do not copy anything — they harvest information content from training data, which gets broken up, and use their learning to generate new text.” In a news article for Nature, Elizabeth Gibney examines the likelihood that published research has been fed into the training-data hopper for generative AI models – whether the author(s) know it or not.
