Aplicação de aprendizado ativo e teoria dos conjuntos Fuzzy em bases de dados não rotuladas na área da saúde

Abstract

Introduction: In recent years, technology has enabled the massive collection and storage of data in the medical field. Artificial Intelligence, and more specifically Machine Learning, has played an important role in offering tools to effectively use this information, facilitating crucial healthcare decisions regarding diagnoses and treatments. However, a persistent challenge is the scarcity of sufficient reliable and labeled data for the effective application of leading machine learning tools. The common scenario involves an abundance of unlabeled information and a scarcity of labeled data, consequences of the complexity and high cost of labeling in healthcare. Objectives: Propose and evaluate new machine learning applications to solve unlabeled data classification problems in health sciences, minimizing the need for extensive labeling. Methods: The approach is quantitative and technological development and produces a new methodology for data processing in the health area. The work began with a bibliographical research focused on creating a machine learning model for an unlabeled database, without the need to develop rules or extensive data labeling. The new proposal included the use of active learning, data grouping and the pre-processing of raw medical data by transformation with fuzzy membership functions. Results: The model was evaluated when reaching 20% of labeled instances. The metrics were accuracy (90.10%), sensitivity (91.20%), specificity (87.36%), precision (94.7%) and AUC (0.91). Conclusion: The results obtained are encouraging, indicating a new interpretable and effective strategy to reduce the amount of labeled data essential in training an underlying model.