Transfer Learning Enhances Neurological Outcome Prediction for Out-of-Hospital Cardiac Arrest: Validation Across Diverse Geographic Contexts.
Studies in health technology and informatics · 2025-08-07 · Method & validation
Abstract
Developing accurate risk stratification models for out-of-hospital cardiac arrest (OHCA) in low-resource settings is challenging due to small sample sizes and poor data quality. The Pan-Asian Resuscitation Outcomes Study (PAROS) network provides a valuable OHCA registry across the Asia-Pacific region, but limited data availability, such as in Vietnam, hampers robust model development. Transfer learning (TL), a machine learning technique, can address these challenges by adapting pre-trained models to new contexts. We used TL to adapt an OHCA neurological outcome prediction model from a large Japanese cohort to Vietnam (n = 243) and Singapore (n = 15,916), representing low- and high-resource settings. TL significantly improved performance in Vietnam (AUROC: 0.807 vs. 0.467) and modestly enhanced outcomes in Singapore, demonstrating its potential to overcome data limitations and improve prediction accuracy in diverse healthcare settings.