Multi-class data augmentation for prediction of postpartum hemorrhage using improved ACGAN

The dataset of primary postpartum hemorrhage (PPH) faces the challenge of insufficient samples, and Generative Adversarial Networks (GANs) have shown considerable promise in addressing the scarcity and imbalance of samples in the diagnosis of PPH. However, existing GAN models often suffer from inherent defects, including mode collapse and gradient vanishing. To surmount these limitations, we propose an innovative supervised model framework, the Multi-Expert Auxiliary Classifier Generative Adversarial Network (MWACGAN), based on an improved Wasserstein distance. Firstly, an independent Deep Neural Network (DNN) classifier is ingeniously integrated to enhance the synergy between discrimination and classification tasks. Secondly, the objective function is designed using the Wasserstein distance with gradient penalty constraints to improve the quality of newly generated sample data and the stability of the training process. The proposed method is employed for the diagnosis of patients with PPH. In comparison to existing algorithms such as Adaptive Synthetic Sampling (ADASYN), traditional Auxiliary Classifier Generative Adversarial Network (ACGAN) and Conditional Generative Adversarial Network (CGAN), etc., this method can generate high-quality PPH sample data more efficiently. It serves as a valuable tool to support the training of deep learning-driven diagnostic models for PPH patients, achieving good stability and high-precision prediction.