After cardiac arrest, families and doctors often face agonizing uncertainty. New Duke-NUS research shows how AI can sharpen predictions and support care in hospitals with limited resources, while global experts push for strong guardrails to keep patients safe.
In hospitals with few specialists and limited equipment, doctors caring for someone after cardiac arrest often must make life-or-death decisions with little more than experience and hope to guide them.
New research led by Duke-NUS Medical School in Singapore suggests artificial intelligence could change that, helping clinicians in resource-limited settings better predict which patients are likely to recover and how best to focus scarce medical resources.
In a study published in npj Digital Medicine, Duke-NUS researchers and collaborators showed that an advanced AI model originally developed in Japan could be successfully adapted for use in Vietnam, where hospitals have far less data and infrastructure.
The team used a technique called transfer learning, which takes a model trained on a very large dataset and fine-tunes it for a new environment with much less local data. That approach is especially promising for low- and middle-income countries, where building powerful AI tools from scratch is often impossible.
The original brain-recovery prediction model was trained on data from 46,918 out-of-hospital cardiac arrest patients in Japan. The Duke-NUS team then adapted it for a Vietnamese hospital setting and tested it on a much smaller group of 243 patients.
Once adapted, the AI tool was able to correctly separate high-risk from low-risk patients about 80% of the time. By contrast, when the original Japanese model was used in Vietnam without adaptation, it was accurate only around 46% of the time.
The findings show that health systems do not need to start from zero to benefit from AI, noted senior author Liu Nan, an associate professor in the Duke-NUS’ Centre for Biomedical Data Science and director of the Duke-NUS AI + Medical Sciences Initiative.
“The study shows AI models to not need to be rebuilt from scratch for every new setting. By adapting existing tools safely and effectively, transfer learning can lower costs, reduce development time and help extend the benefits of AI to healthcare systems with fewer resources,” Liu said in a news release.
For families and clinicians, more accurate predictions after cardiac arrest can help guide difficult conversations about treatment options, rehabilitation and long-term care. For hospitals with limited beds, staff and equipment, better risk prediction can also support fairer and more efficient use of resources.
Advancing AI in Low-Resource Settings
Beyond this single example, Duke-NUS researchers say AI could support many aspects of care in low-resource environments, from diagnosis to triage to clinical decision-making.
In a separate study published in Nature Health, Duke-NUS scientists and collaborators, including colleagues from University College London (UCL), examined how large language models, or LLMs, could advance global health. LLMs are AI systems trained on massive amounts of text so they can understand and generate human language.
The researchers highlighted real-world examples already in use. In South Africa, a chatbot is providing pregnancy-related information to expecting mothers, helping to fill gaps where in-person care is hard to access. In Sierra Leone, community health workers are using smartphone-based tools to detect malaria infections from blood smear images, offering a cheaper alternative to traditional microscope-based systems.
Yet the team found that most AI development and deployment still happens in high-income and upper-middle-income countries. Many low- and middle-income nations face basic barriers such as unreliable internet and electricity, limited computing infrastructure, a shortage of trained AI specialists and little local guidance on how to close these gaps.
Co-author Siegfried Wagner, from the UCL Institute of Ophthalmology and Moorfields Eye Hospital NHS Foundation Trust, emphasized that the greatest need is often where AI could have the greatest impact.
“LLMs have the greatest opportunity to transform healthcare in settings where specialist physicians are scarcest, but the global health community needs to work together with some urgency to ensure the implementation of LLMs is supported in regions where adoption is most challenging,” he said in the news release.
The Duke-NUS team argues that technology alone is not enough. To make AI tools truly useful and safe, health workers and communities must be able to understand and trust them.
Ning Yilin, a senior research fellow at Duke-NUS’ Centre for Biomedical Data Science and a co-first author of the Nature Health study, emphasized that building skills and confidence among frontline staff is essential.
“Strengthening digital literacy and building confidence in using these tools will ensure AI supports, rather than disrupts, the workforce,” she said. “Tailored skills-development pathways can help under-resourced workers adapt and thrive, allowing AI to uplift and add value to clinical and administrative roles.”
Building global consensus for AI guardrails
At the same time, the rapid spread of AI in medicine is raising urgent questions about ethics, safety and accountability. Traditional medical device regulations often do not address AI-specific risks, such as how to handle patient data, what to do when models generate incorrect or misleading outputs, or who is responsible when an AI-assisted decision harms a patient.
To tackle these challenges, Duke-NUS researchers and partners have proposed an international consortium called the Partnership for Oversight, Leadership, and Accountability in Regulating Intelligent Systems–Generative Models in Medicine, or POLARIS-GM.
The proposed group would bring together health care leaders, regulators, ethicists and patient representatives from around the world to develop practical guidance for governing AI in medicine. Its goals include setting best practices for evaluating new tools, monitoring their real-world impact, establishing safety guardrails and ensuring that regulations work for both high-resource and low-resource settings.
POLARIS-GM plans to start by reviewing existing research and regulatory approaches, then work toward a global consensus on how to oversee AI in health care.
Jasmine Ong, from the Duke-NUS AI + Medical Sciences Initiative and a principal clinical pharmacist at Singapore General Hospital, is first author of a correspondence on this effort published in Nature Medicine.
“With clear oversight and clearly defined guidelines, healthcare systems can confidently leverage AI’s many strengths to improve health outcomes while steering clear of potential pitfalls. From policymakers to patient groups, all stakeholders have a crucial role to play in making this goal a reality,” she added.
Taken together, the Duke-NUS studies point to a future in which AI can help close some of the most stubborn gaps in global health. By adapting powerful models for new settings, supporting frontline workers and building strong governance, the researchers say AI could help ensure that where a person lives or how much a hospital can afford matters less in determining the care they receive.
For patients and families facing the uncertainty of a cardiac arrest or other serious illness, that could mean clearer answers, more consistent care and, in some cases, a better chance at recovery.
Source: Duke-NUS Medical School