UNRAVELING LEARNING CONFLICTS IN SUPERVISED LEARNING DATASETS: A METHODOLOGICAL APPROACH FOR IMPROVED MACHINE LEARNING PERFORMANCE

Abstract

The domain of image analysis utilizing artificial intelligence has expanded significantly due to advancements in neural networks. A particularly promising domain is medical diagnosis via lung X-rays, essential for identifying diseases such as pneumonia, which may be confused with other ailments. Notwithstanding medical expertise, accurate diagnosis remains difficult, and this is where proficient algorithms can provide assistance. Nonetheless, the analysis of medical images poses difficulties, particularly when datasets are constrained and imbalanced. While strategies to balance these classes have been investigated, there remains a deficiency in understanding their local impact and influence on model evaluation. This study seeks to examine the impact of class imbalance in a dataset on the efficacy of metrics employed to assess predictions. It illustrates that class separation within a dataset influences trained models and warrants greater consideration in future research. Classification models utilizing artificial and deep neural networks are developed in the R environment to attain these objectives. These models are trained utilizing a collection of publicly accessible images pertaining to lung pathologies. All outcomes are corroborated using metrics derived from the confusion matrix to assess the influence of data imbalance on the efficacy of medical diagnostic models. The findings prompt inquiries regarding the methodologies employed to categorize classes in numerous studies, with the objective of attaining class equilibrium in imbalanced datasets, and suggest new directions for future research to explore the effects of class delineation in datasets associated with clinical pathologies.
Keywords: ; ;; ;