AI Model Accurately Predicts Neurological Recovery After Cardiac Arrest

This challenge is even greater in hospitals with limited access to advanced diagnostics, large datasets, and specialist expertise. Similar barriers affect many areas of healthcare in low- and middle-income countries, where delayed or uncertain decisions can worsen outcomes. Now, new research shows that artificial intelligence (AI) can help reduce this uncertainty by improving diagnostic accuracy and patient outcome prediction, even in resource-constrained settings.

Researchers from Duke-NUS Medical School (Singapore) adapted an advanced AI model to predict neurological recovery after cardiac arrest using a technique called transfer learning. Transfer learning allows AI models trained on large datasets in high-resource settings to be safely adapted for use in new environments with limited local data. Instead of building new models from scratch, existing algorithms are fine-tuned to reflect local patient populations, reducing cost, development time, and data requirements.

The researchers adapted a brain-recovery prediction model originally developed in Japan using data from 46,918 out-of-hospital cardiac arrest patients. The model was then tested in Vietnam on a much smaller cohort of 243 patients. The findings, published in npj Digital Medicine, show that when applied directly, the original model performed poorly, but after transfer learning, it correctly distinguished high-risk from low-risk patients about 80 percent of the time, compared with around 46 percent previously.

In addition to cardiac care, AI tools show promise across a wide range of applications in resource-limited healthcare systems. In a separate study, the Duke-NUS researchers and collaborators explored how large language models could support diagnostics, clinical decision-making, and access to care in low-resource settings. Examples include chatbots providing pregnancy guidance in South Africa and smartphone-based tools helping community health workers detect malaria in Sierra Leone. These approaches could help bridge gaps where specialist care and laboratory infrastructure are scarce.

Despite growing potential, AI adoption remains concentrated in high-income regions, with many countries facing barriers such as limited infrastructure, lack of expertise, and weak regulatory frameworks. Existing medical regulations often fail to address AI-specific risks, including data privacy, bias, and unreliable model outputs. To address these challenges, researchers have proposed an international consortium to guide responsible AI governance, focusing on safety standards, accountability, and adaptation for resource-limited environments.

“The study shows AI models to not need to be rebuilt from scratch for every new setting,” said Associate Professor Liu Nan, senior author of the study. “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.”

Related Links:
Duke-NUS Medical School