MGHNM: A multi-granularity based on hybrid network model for postpartum hemorrhage prediction

Precise prediction of postpartum hemorrhage (PPH) is of great significance for early identification of high-risk pregnant women, optimizing medical resource allocation, and reducing maternal mortality. However, existing PPH prediction methods suffer from limitations such as coarse prediction granularity, and single-stage prediction processes, leading to insufficient prediction accuracy. This has made prediction methods based on hybrid network architectures an important research direction in current PPH studies. This paper proposes a Multi-Granularity Hybrid Network Model (MGHNM) for PPH prediction, which integrates advanced methods such as ensemble learning, convolutional neural networks (CNN), and variational autoencoders (VAE). By leveraging multi-level feature extraction, the model effectively suppresses interference from secondary information, thus significantly enhancing prediction accuracy. The MGHNM model introduces a learnable control switch mechanism to achieve dynamic feature selection, significantly enhancing the model's discriminative ability. By organically combining the CatBoost classifier, CNN feature extractor, VAE representation learning module, and Vision Transformer (ViT), the hybrid network prediction model achieves a significant improvement in prediction accuracy for the three-level classification task of PPH severity (mild, moderate, and severe). The experimental data in this paper is derived from a PPH dataset constructed from the electronic medical record (EMR) system of the Maternal and Child Health Hospital in Jinan, Shandong Province, China. Three experiments were designed: First, the hyperparameters of the prediction model were optimized and analyzed. Second, a multi-model comparative experiment was conducted. Finally, an ablation study was performed. The experimental results demonstrate the significant superiority of the proposed MGHNM model for PPH prediction. It achieves an overall mean accuracy of 89.50 % with a standard deviation of 0.0045 %, substantially outperforming both the baseline and state-of-the-art (SOTA) methods.