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.
October 24, 2025
In this week’s Duke AI Health Friday Roundup: mimicking brain “modularity” for better AI models; bumblebee empathy; “fragility” of GPT-5 output in medical settings; mRNA COVID vaccines boost immune response against some cancers; human vs LLM in medical diagnosis; framework for evaluating risk of memorization in EHR-trained foundation models; AI-generated lesson plans fail to impress; much more:
AI, STATISTICS & DATA SCIENCE
- “…we propose Mixture of Cognitive Reasoners (MiCRo): a modular, transformer-based architecture post-trained with a curriculum that induces functional specialization across experts. Concretely, we partition the layers of a pretrained language model into four expert modules aligned with well-studied cognitive networks in the human brain. MiCRo offers three key advantages over standard language models…Taken together, cognitively grounded functional specialization yields models that are both more human-like and more human-interpretable.” A research article by Alkhamissi and colleagues, available as a preprint at arXiv, presents a modular approach to machine reasoning with specialized components.
- “For medicine, these vulnerabilities have direct consequences. A hospital-approved deployment might include rules such as ‘Never provide drug dosages without citing an approved guideline’, or ‘Always direct emergent chest pain queries to call emergency services’. If instruction hierarchy breaks down, an adversarial patient prompt or even an innocently ambiguous clinician query could bypass those safeguards…” A commentary published in Nature Medicine by Handler and colleagues spotlights the “fragility” of the GPT-5 large language model when employed in healthcare settings.
- “…it’s not so much that the models have adopted our goals, or even pretend to, as an actor might. They just learn, statistically, patterns of text that describe common goals, reasoning steps and behaviour, and they regurgitate versions of them. The problem will probably get worse once future LLMs train on papers describing LLM scheming, which is why some papers elide some details.” A Nature feature article by Matthew Hutson scrutinizes studies that revealed “scheming” behavior by large language models, in which the AIs exhibit unsavory or alarming behavior.
- “High-profile breaches, such as Stanford Hospital’s 20M lawsuit over leaked records, underscore the urgency of systematically auditing AI models in healthcare. To address this, we introduce an evaluation framework for EHR-FM that measures memorization and its associated privacy risks. Our tests quantify different forms of memorization and assess their implications in clinical settings, distinguishing harmful leakage at the patient level from useful generalization at the population level.” A research article by Fallahpour and Ghassemi, available as a preprint from arXiv, proposes a framework for assessing the risk of memorization with AI foundation models trained on EHR data.
BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH
- “Affective contagion, a core component of empathy, has been widely characterized in social vertebrates but its existence in any invertebrate is unknown. Using a cognitive bias paradigm we demonstrate positive affective contagion in bumble bees….Our findings suggest that affective contagion may be an evolutionarily widespread mechanism present in both social vertebrates and social insects.” Bees got feelings: A new study published in Science by Romero-González and colleagues presents evidence that bumble bees exhibit affective contagion – a key component of empathy in other animals.
- “The data at 12 months in this clinical study showed that the PRIMA subretinal implant restored meaningful central vision in persons with geographic atrophy due to AMD, thus enabling the performance of visual tasks such as reading and writing. Although no retinal implants have been explanted in humans, the wireless design allows for replacement with higher resolution next-generation implants…” A research article published by Holz and colleagues in the New England Journal of Medicine describes the successful use of an implant to restore vision loss in persons with age-related macular degeneration.
- “…researchers analysed the medical records of more than 1,000 people with lung cancer or melanoma. They found that, in people with a certain type of lung cancer, receiving an mRNA COVID-19 vaccine was linked to a near doubling in survival time, from 21 months to 37 months. Unvaccinated people with metastatic melanoma survived an average of 27 months; by the time data collection ended, vaccinated people had survived so long that the researchers couldn’t calculate an average survival time.” Nature’s Max Kozlov reports on new findings from both human and animal studies that suggest mRNA COVID vaccines may help boost the effectiveness of checkpoint inhibitor therapies for cancer.
COMMUNICATIONS & Policy
- “Overall, we found the AI-generated lesson plans to be decidedly boring, traditional, and uninspiring. If civics teachers used these AI-generated lesson plans as is, students would miss out on active, engaged learning opportunities to build their understanding of democracy and what it means to be a citizen….Our study emphasizes the need for teachers to be critical users, rather than quick adopters, of AI-generated lessons. AI is not an all-in-one solution designed to address the needs of teachers and students.” In an article co-published by Ars Technica and The Conversation, Torrey Trust and Robert Maloy deliver a critical assessment of the use of AI utilities in educational settings, particularly in developing lessons plans for students.
- “For the time being, embracing the uncertainty inherent in diagnosis with humility, practicing careful clinical observation, continually calibrating and refining our knowledge of diseases, and mastering the art of history taking will hopefully let us keep our edge (for a little while longer). While I don’t know that I would like to step in the ring again with a newer and more powerful LLM, I remain hopeful that these tools will augment, not replace us.” In a perspective article for NEJM AI, Daniel Restrepo offers a perspective on the relative merits of human vs AI-based diagnosis after going head-to-head with GPT-4 in a Grand Rounds session at Harvard.
- “…how can regulators ensure that algorithms set fair prices? Their traditional approach won’t work, as it relies on finding explicit collusion. ‘The algorithms definitely are not having drinks with each other,’ said Aaron Roth, a computer scientist at the University of Pennsylvania…Yet a widely cited 2019 paper showed that algorithms could learn to collude tacitly, even when they weren’t programmed to do so….The end result was high prices, backed up by mutual threat of a price war.” Quanta’s Ben Brubaker looks at the theoretical foundations of “collusion” by pricing algorithms that can drive price increases.
