Geração de dados sintéticos em bases médicas com amostras reduzidas para a predição de câncer

Abstract

During the application of Artificial Intelligence techniques in healthcare, obtaining large clinical datasets remains a challenge due to data collection costs, limited patient availability, and ethical constraints. This limitation is particularly relevant in studies based on laboratory tests, biochemical markers, and enzymatic assays, potentially affecting the training and generalization of machine learning models. As an alternative, synthetic data generation enables artificial expansion of medical datasets. This work presents an experimental analysis of synthetic data generation methods applied to cancer prediction in medical datasets with limited sample sizes. For this purpose, the Tabular Prior-Fitted Generator (TabPFGN) and Generative Adversarial Networks (GANs) are compared using the Tabular Prior-Data Fitted Network (TabPFN) and Extreme Gradient Boosting (XGBoost) models. The results demonstrate that the use of synthetic data can improve classifier performance, achieving gains of up to 6.6% in F1-score compared to the baseline scenario, with statistically significant improvements observed in part of the experiments.