This study demonstrates the feasibility and utility of TL in adapting existing predictive models to new clinical contexts, especially in LDRS. Using OHCA as a proof of concept, we showed that TL can significantly enhance model performance with smaller datasets, as observed in Vietnam, and provide incremental benefits with larger datasets, as seen in Singapore. These findings highlight TL’s broad applicability across diverse global healthcare settings, offering a scalable solution for developing robust predictive models and reducing healthcare disparities by tailoring models to specific regional contexts.

Healthcare research in LDRS faces numerous challenges, including small population sizes, low outcome prevalence, and disparities in epidemiological features and data collection systems. Importantly, even high-income regions are not immune to these issues. For example, in OHCA, where favorable neurological outcomes are rare (3–4%), large datasets are essential for robust model development. The external model used in this study was developed from 46,918 OHCA patients in Japan over 5–6 years²¹, providing a solid foundation for model development and validation. In contrast, Singapore, with only approximately 3,000 annual OHCA cases, would require more than a decade to accumulate a comparable dataset. While both Singapore and Japan are considered developed countries, conducting such research in Singapore is more challenging due to differences in population size. The situation is even more difficult in data-constrained settings like Vietnam, where OHCA databases are still emerging. Despite approximately 4,000 annual cases, it would still take over 10 years to collect enough data for robust model development. These prolonged timelines delay the implementation of predictive models that could improve patient outcomes and exacerbate global healthcare disparities, highlighting the urgent need for innovative solutions to address data gaps in LDRS.

AI techniques like TL provide an effective solution by adapting existing models to local populations, improving predictive performance even with scarce data. This study highlights the transformative potential of TL for developing accurate prediction models in LDRS, where socioecological factors, in addition to financial constraints³, influence healthcare outcomes. In Vietnam (n = 243), TL significantly enhanced model accuracy. In contrast, in Singapore (n = 15,916), TL provided more modest gains, illustrating its adaptability across different resource levels. Despite Singapore’s larger dataset, the annual OHCA case numbers remain insufficient to independently develop models comparable to those built with Japan’s extensive data. These findings underscore the utility of TL in leveraging external information to overcome local dataset limitations, enabling the development of clinically relevant models tailored to specific populations. By addressing heterogeneity in data distributions and clinical practices, TL enhances the reliability and applicability of predictive models, making AI technologies more accessible and impactful in data-limited settings.

As shown in Table 2, the benefits of TL were most pronounced with the Vietnam cohort, where the improvement in overall prediction performance was more significant than in Singapore. This supports the theoretical foundations of TL¹⁰, which is particularly effective when the source data (e.g., the Japanese cohort) has a large sample size and the target data (e.g., Vietnam and Singapore) is relatively limited. This study demonstrates the applicability of TL in challenging settings like LDRS, showing how researchers can directly leverage knowledge from external studies without requiring access to additional datasets or direct collaborations. Federated learning (FL), another relevant AI technique, addresses data privacy concerns in cross-site collaboration by enabling co-training of AI models without sharing data²⁶. Unlike FL, which requires simultaneous participation from multiple data owners, TL allows independent adaptation of publicly available models. This makes TL particularly advantageous for researchers with limited access to external data or collaboration opportunities, offering a practical way to overcome barriers associated with data sharing and resource constraints¹⁰.

TL is a flexible AI approach applicable to a wide range of models and knowledge transfer scenarios¹⁰, such as transferring insights on drug sensitivity between different cancer types²⁷ or adapting knowledge of surgical complications across patient groups²⁸. In LDRS, where collecting large, high-quality datasets is challenging, TL offers an efficient way to leverage existing models for improved predictive performance. Regardless of the clinical or biomedical domain, as Li et al.¹⁰ highlight, the successful implementation of TL depends on carefully selecting an appropriate external study and ensuring that the chosen TL framework aligns with the research question. It is also crucial to address potential privacy concerns when building or using external models. When co-training is required, FL can be integrated into TL frameworks to enable secure collaborations across multiple sites, ensuring data privacy while still facilitating knowledge transfer.

One reason for TL’s effectiveness in this study may be the standardized data format and consistent variable definitions under the PAROS study framework. The Utstein data format, adopted in this study, has been globally accepted since its establishment in 1995, and most countries utilize it for OHCA data collection²⁹. Furthermore, the prehospital management pathways for OHCA cases follow similar protocols across the study countries—paramedics respond to an emergency call, perform resuscitation, and transport patients to hospitals. During this clinical pathway, prehospital OHCA data are systematically collected, which likely facilitated effective knowledge transfer using TL.

Another potential reason is that certain predictors, such as ROSC status, are expected to be strongly associated with patient outcomes, despite variability in unmeasured clinical factors like in-hospital care or hospital systems among the three countries³⁰. Given these considerations, further research is essential to investigate the applicability of TL to the other clinical scenarios where unmeasured factors may significantly impact outcomes. For example, in patients with refractory VF, treatment strategies vary markedly across countries. In Japan, advanced invasive resuscitation procedures, such as extracorporeal cardiopulmonary resuscitation, are commonly performed and may contribute improved outcomes³⁰. In contrast, such procedures are rarely utilized in Singapore and Vietnam³⁰. It remains unclear whether TL would perform well in contexts where critical unmeasured factors, such as the treatment strategy or in-hospital care, vary substantially and may significantly affect patients’ outcomes.

Future research should explore strategies for handling ultra-small datasets, where outcome prevalence is extremely low or absent. In this study, the Vietnam dataset included 234 cases with only 13 positive outcomes, but settings with even fewer cases present greater challenges. In such instances, the TL method used here may not be directly applicable. To enhance model robustness and improve predictive reliability, alternative approaches—such as synthetic data generation and oversampling—should be integrated with TL. Combining these techniques within TL frameworks may further mitigate data scarcity in low-resource clinical settings.

This study has several limitations. First, while the Singapore data were derived from a nationwide population-based registry for OHCA patients and are likely representative of the Singapore population, the Vietnam data were collected from a limited number of hospitals in Ho Chi Minh, Hanoi and Hue. As a result, the Vietnam data may not be representative of the broader population, potentially limiting the generalizability of the model to current clinical settings in Vietnam. Second, although the data from Vietnam, Singapore, and Japan were collected using an internationally standardized format with consistent definition, differences in local clinical practices and resuscitation protocols may have influenced data collection and reporting. This might lead to a risk of measurement bias. Third, prospective external validation using independent data would be ideal for confirming the robustness of our results. However, at present, no such dataset is available in Vietnam or Singapore for this process. Further research is warranted to assess reproducibility and validate these findings in independent cohorts.

In conclusion, this study demonstrates that TL has the potential to bridge regional resource disparities in healthcare. While applied here to OHCA, TL’s adaptability extends to other emergency conditions with low event rates, such as trauma, sepsis, heart failure, and stroke. Moreover, TL holds broader potential beyond emergency medicine, offering a scalable, efficient strategy to enhance clinical decision-making in LDRS globally. By reducing reliance on large-scale data collection, TL facilitates equitable global access to AI-driven predictive models, helping to reduce healthcare disparities and optimize patient outcomes in data-constrained environments.