Transformer-based Self-Supervised Imputation and Attention GANs Oversampling for Medical Data Processing

Dealing with missing data and class imbalance is a necessity for the development of a solid predictive model in medical research. This work presents a novel hybrid dynamic method devoted to addressing these dual challenges, specifically focusing on medical datasets. In this work, we investigate the incompleteness and imbalance of the dataset using two sophisticated measures: the Incompleteness Factor and the Imbalance Ratio. These metrics are used to guide customized approaches. A transformer-based imputation method is employed to learn complex data dependencies to impute missing data. Second, the oversampling method of Self-Attention Enhanced Conditional Generative Adversarial Network is implemented. This technique generates better model training results by accommodating selfattention mechanisms that create fake samples that follow the actual data distributions as well as class distributions. This paper introduces one potential solution to obtain model-level accuracy increase for tasks like disease prediction, patient prognosis, etc., over medical datasets. The proposed method aims to remove the measurement biases caused by incompleteness and class imbalance and preprocess the data to enhance the credibility and clinical relevance of the experimental results.