Avaliação do modelo TabPFN na predição de diabetes com variação no número de instâncias

Abstract

Diabetes mellitus is one of the most common chronic diseases worldwide, representing one of the greatest public health challenges of today. This condition affects the body’s ability to produce or properly use insulin, the hormone responsible for regulating blood glucose levels, contributing to the development of secondary complications and highlighting the importance of early diagnosis. With advances in machine learning techniques, these approaches have been increasingly used to support the prediction of diseases such as diabetes; however, the scarcity of clinical data, linked to data protection policies and resource limitations, still represents a significant obstacle. In this context, the TabPFN model stands out as a Transformer-based network capable of performing fast and accurate predictions even with limited amounts of data. Given this scenario, this work evaluates the performance of TabPFN in predicting cases of diabetes mellitus, analyzing its behavior under different data volumes and comparing it with well-established models in the literature. The results demonstrated that TabPFN achieved highly competitive performance, standing out from 100 training instances onward and reaching an F1-score close to 0.80 with 1,000 samples, even without the use of balancing techniques, while also identifying glucose-related variables as the most decisive for classification. It is concluded that TabPFN represents a promising alternative for diabetes prediction in data-scarce scenarios.