JMIR Publications

ABSTRACT

Hemodialysis is crucial for patients with end-stage renal disease; however, intradialytic hypotension (IDH) is a relatively common complication of hemodialysis, with serious consequences. Existing machine-learning-based IDH prediction models have limitations, including not taking account of individual differences and key clinical elements, and having high data-quality requirements. To address these issues, this study aimed to develop an Integrated Multimodal Clustering-Augmentation (IMCA) model to predict IDH. The model integrates daily measurement data and dialysis machine time-series data, and labels IDH after excluding drug-induced blood pressure changes. The model consists of an electronic health record clustering module that analyzes and integrates patient data to extract effective information; a data-augmentation module using SMOTE-TomekLink and Wasserstein GAN-GP to learn features and balance samples; and a comprehensive prediction module that selects different models for targeted prediction within different clusters according to various data. Experiments using a HD dataset of 412 patients showed that the IMCA-IDH model was robust and accurate, even with missing and imbalanced data. Different model combinations achieved excellent performances in various patient clusters, with a maximum F1 score of 0.989 and an area under the curve of 0.745. This was a single-center retrospective analysis, and further multi-center studies are needed to improve the model’s generalizability. Nevertheless, this model provides a new approach for IDH prediction and references for medical data modeling.