MI-Net: a dual-stream feature network based on Mamba-convolutional neural network (CNN) fusion to predict intradialytic hypotension using multimodal data.

Background: Intradialytic hypotension (IDH) is a frequent and severe complication in hemodialysis patients. Traditional prediction models rely on single-modal data or simple machine learning methods, while Transformer-based deep learning approaches suffer from high computational complexity and poor efficiency in processing long sequential data. Therefore, this study aimed to develop a stable, efficient, and generalizable deep learning model for the early prediction of IDH. Methods: We proposed MI-Net, a dual stream feature network fusing Mamba and convolutional neural network (CNN) for IDH prediction using multimodal data (echocardiography videos and clinical parameters). The CNN branch extracted local spatial features, and the Mamba branch captured long range temporal dependencies. A cross attention fusion module was used to integrate imaging and clinical features. The model was validated with internal data from the Central Hospital of Enshi Tujia and Miao Autonomous Prefecture (CHETM) and external data from Renmin Hospital of Wuhan University (RHWU). Results: On the CHETM test set, MI Net achieved an accuracy of 0.924 and an area under the curve (AUC) of 0.952. On the external RHWU test set, it obtained an accuracy of 0.894 and an AUC of 0.851. Ablation studies confirmed that Mamba, cross attention, and dual stream structure significantly improved performance. Conclusions: MI Net efficiently combines Mamba and CNN, showing high accuracy, strong specificity, and good generalizability. It provides a reliable intelligent tool for early prediction of IDH.