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 28, 2026
In this week’s Duke AI Health Friday Roundup: what happens when AI alone beats human-in-the-loop; COVID vaccination associated with lower risk of myocarditis; study reports widespread use of LLMs in biomedical literature; agentic AI in medicine and a new digital divide; more cases of bogus validation data for antibody manufacturers alleged; FDA approves new KRAS therapy for pancreatic cancer; much more:
AI, STATISTICS & DATA SCIENCE
- “Data from medicine and other fields suggest that when AI-alone performance is consistently superior to human-alone performance, AI alone surpasses human-AI hybrids. Paradoxically, hybrid care in which humans are in (or on) the loop to correct AI errors is likely to worsen rather than improve AI performance…medicine is rapidly approaching the transition point at which AI alone will exceed physicians and physician-AI hybrids in providing the best care at 5 fundamental cognitive medical tasks.” A thought-provoking perspective by Emanuel and colleagues, published in JAMA, weighs the question of whether preserving human-in-the-loop AI in some applications could actually lead to worse performance than allowing the AI to proceed by itself. Over at his Substack, Robert Wachter (whose book is cited by the authors) offers a riposte.
- “…we develop a continuous-time Markov chain formulation for household transmission dynamics in which the model state space is expanded to include transmission graphs describing infection direction and order, with idealised pathogen genomic data used to identify the transmission histories compatible with observations. We conduct simulation studies which show that incorporating genetic information substantially concentrates the regions of high likelihood compared with models based on epidemiological data alone.” A preprint article by Sayyar and colleagues, available from arXiv, presents a new inference model that includes both epidemiological and genomic data in predicting disease transmission at the household level.
- “We argue that access to agentic research capacity should be treated as a shared scientific resource and potentially coordinated at national and international levels, not left to local advantage. The key task is not only to improve agent performance, but to shape the conditions under which agentic science remains trustworthy, plural, and broadly accessible, enabling shared scientific freedom.” A perspective article published this month in NEJM AI by Masanneck and colleagues examines the question of whether the adoption of agentic AI in biomedical research risks creating a “digital divide” between systems and institutions with greater resources and those with less.
- “…we examined how physicians edit AI-generated replies and how different types of edits relate to response time. The findings show that editing burden is not uniformly distributed: Edits requiring clinical judgment impose the greatest per-message time burden, while high-frequency administrative edits contributed the greatest cumulative workload across the health system.” A research article by Poursoltan and colleagues, published in NEJM AI, presents results from a study evaluating physician’s editing practices when working with LLM-generated patient messages.
BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH
- “In this large cohort of adolescents and young adults aged 16–25 years, SARS-CoV-2 infection was associated with higher 6-month risks of myocarditis, pericarditis, and all-cause mortality than BNT162b2 vaccination. The same direction of association was consistent across competing-risk, propensity-matched, and Cox analyses and aligns with international cohort studies and meta-analyses. Together they support a favorable observed benefit-risk profile for vaccination compared with infection.” A research article published in the journal Vaccine by Toraih and colleagues explores the issue of myocarditis in younger persons who were vaccinated for COVID vs those who were infected by the virus (H/T UMN CIDRAP).
- “Through a centralized referral network, supported by genetic counseling and multidisciplinary teams, and close clinical–laboratory coordination, the program achieved seamless integration into acute-care workflows, demonstrated robust diagnostic performance and delivered measurable clinical impact in a population characterized by high consanguinity and a considerable burden of autosomal recessive disease.” A research article by Rabea and colleagues, published in Nature Medicine, describes the rollout of a rapid whole-genome sequencing program for pediatric patients with critical illnesses at NICU and PICU facilities across Dubai.
- “This update is to say that this newfound distrust should probably extend to a good fraction of the research antibody industry at large. Our repository now contains 18,943 images presented as validation data for 17,495 unique antibody products sold by 15 different vendors.” At a blog focused on research integrity issues, Reese Richardson portrays the scope of unfolding evidence suggesting that multiple companies, including some of those most prominent in the biomedical industry, have been selling antibodies whose performance claims are buttressed by fabricated validation data.
- “The Food and Drug Administration on Wednesday approved a life-extending treatment for advanced pancreatic cancer…that is the first to attack a genetic cause of the aggressive, highly lethal malignancy….The FDA’s approval was supported by a practice-changing clinical trial that read out earlier this year. Patients with advanced pancreatic cancer who received Rasonque as a second-line treatment achieved a median overall survival of 13.2 months, compared to 6.7 months for patients offered standard chemotherapy.” STAT News’ Adam Feuerstein and Angus Chen report on this week’s decision by FDA to approve the drug daraxonrasib (Rasonque) for the treatment of advanced pancreatic cancer.
COMMUNICATIONS & Policy
- “AI presents substantial practical concerns for our community, from data privacy and confidentiality to questions around disparate access, bias, fairness, and accountability. AI’s explosive growth and ascendance also raise important questions for society, from the environmental impact of AI data centers, to the use of intellectual property and training data, to the overall human impact of the technology and the industry…” Earlier this August, an ad-hoc committee at MIT released a report on the use of AI in instruction, learning, and training at the university.
- “…we suggest and validate a new unbiased approach to estimate LLM usage in a corpus of texts based on changing word frequencies. We apply our method to the full texts of open-access biomedical papers from Pubmed Central, and show that by the end of 2025, 89% of papers show excess of LLM-associated vocabulary.” A preprint article by Holzwarth and colleagues, available at arXiv, presents evidence from a study that surveyed the corpus of biomedical literature available on PubMedCentral and published in 2025 for evidence of AI-assisted writing.
