Cardiac disease diagnosis based on GAN in case of missing data

Abstract

AbstractIn daily life, two common algorithms are used for collecting medical disease data: data integration of medical institutions and questionnaires. However, these statistical methods require collecting data from the entire research area, which consumes a significant amount of manpower and material resources. Additionally, data integration is difficult and poses privacy protection challenges, resulting in a large number of missing data in the dataset. The presence of incomplete data significantly reduces the quality of the published data, hindering the timely analysis of data and the generation of reliable knowledge by epidemiologists, public health authorities, and researchers.Consequently, this affects the downstream tasks that rely on this data. To address the issue of discrete missing data in cardiac disease, this paper proposes the AGAN (Attribute Generative Adversarial Nets) architecture for missing data filling, based on generative adversarial networks. This algorithm takes advantage of the strong learning ability of generative adversarial networks. Given the ambiguous meaning of filling data in other network structures, the attribute matrix is designed to directly convert it into the corresponding data type, making the actual meaning of the filling data more evident. Furthermore, the distribution deviation between the generated data and the real data is integrated into the loss function of the generative adversarial networks, improving their training stability and ensuring consistency between the generated data and the real data distribution. This approach establishes the missing data filling mechanism based on the generative adversarial networks, which ensures the rationality of the data distribution while filling the missing data samples. The experimental results demonstrate that compared to other filling algorithms, the data matrix filled by the proposed algorithm in this paper has more evident practical significance, fewer errors, and higher accuracy in downstream classification prediction.