Transfer learning and governance help bridge healthcare AI divide

Researchers in Singapore have shown that advanced artificial intelligence (AI) techniques can significantly improve clinical diagnostics in countries with limited resources without the need for massive local datasets.

A team from Duke-NUS Medical School has successfully applied transfer learning, a method where a model developed for one task is reused as the starting point for another, to predict patient outcomes after cardiac arrest.

The study, published in npj Digital Medicine, addresses a common challenge in AI adoption in low- and middle-income countries, which is the lack of extensive, high-quality data required to train algorithmic models from scratch.

To test the effectiveness of transfer learning, the researchers used a brain-recovery prediction model originally built in Japan using data from 46,918 out-of-hospital cardiac arrest patients. They adapted this model for use in Vietnam, testing it on a smaller group of 243 patients.

The results showed a huge improvement in diagnostic accuracy. When the original Japanese model was applied directly to the Vietnamese context, it distinguished high-risk from low-risk patients with 46% accuracy. However, the adapted transfer learning model achieved an accuracy rate of around 80%.

“The study shows AI models do not need to be rebuilt from scratch for every new setting,” said Liu Nan, associate professor at Duke-NUS’s Centre for Biomedical Data Science. “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.”

Despite the growing potential of AI in healthcare, adoption of the technology remains uneven across the globe. In a separate study published in Nature Health, Duke-NUS researchers and collaborators such as University College London (UCL) noted that while 63% of surveyed healthcare providers use AI tools, adoption is more prevalent in high- and upper-middle-income countries.

The research highlighted the potential for large language models (LLMs) to improve access to care, diagnostics and clinical decision-making in low- and middle-income countries that continue to face adoption barriers such as limited infrastructure and expertise.

Examples include Sierra Leone, where community healthcare workers use smartphone apps to detect malaria infections from blood smear samples, a more cost-efficient method than conventional microscope-based systems. And in South Africa, chatbots provide pregnant mothers with prenatal advice.

“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,” said Siegfried Wagner from UCL Institute of Ophthalmology and Moorfields Eye Hospital NHS Foundation Trust.

Ning Yilin, senior research fellow at the Centre for Biomedical Data Science at Duke-NUS, added that empowering people should be the priority when integrating LLMs into healthcare.

“Strengthening digital literacy and building confidence in using these tools will ensure AI supports, rather than disrupts, the workforce. Tailored skills-development pathways can help under-resourced workers adapt and thrive, allowing AI to uplift and add value to clinical and administrative roles,” she said.