The use of machine learning (ML) in healthcare has expanded rapidly, as these algorithms can learn patterns from data to support critical decision-making, such as predicting patient outcomes[1]. In developing countries, challenges such as limited diagnostic resources, regional disparities, and socioeconomic inequalities hinder data collection, highlighting the need for innovative solutions like transfer learning.

One potential approach is to develop predictive algorithms in locations with large data availability and resources and then investigate whether these algorithms can be generalized and adapted in other settings without requiring major modifications. Transfer learning presents a promising solution in these cases, as it leverages knowledge acquired from a source domain to initialize and accelerate model training in a target domain [2], [3]. This approach can be especially valuable in areas where data collection is costly or challenging, leading to limited sample sizes.

Transfer learning is frequently applied in image recognition and classification tasks using deep neural networks. A common approach involves leveraging a pre-trained model on a large dataset of images to identify general features, such as lines, corners, and edges. This model can then be fine-tuned for the specific task of disease detection in medical images, enabling more efficient and accurate results [4], [5], [6]. While image classification is the most common application of transfer learning, this approach can also be applied to various other contexts. For instance, a study by Lee and colleagues [7] used a regression-based model to transfer valuable information from a dominant racial group (white) to enhance the prediction of a measure related to visual function in other racial groups, such as African American and Asian populations. Hammour and colleagues [8] also used a different approach of transfer learning technique based on decision trees to improve the accessibility of sleep monitoring in older adults.

Our study aims to apply machine learning methods to predict Intensive Care Unit (ICU) admissions for COVID-19 patients within a Brazilian multicenter setting. By leveraging data from multiple hospitals across diverse regions, the study captures Brazil’s large socioeconomic and resource disparities, as well as varying sample sizes. Specifically, we will evaluate each hospital's predictive performance, identify the best-performing hospital, validate its generalizability across other hospitals through external validation, and determine whether transfer learning models can further improve outcomes.