Multi-Channel Hypergraph-Enhanced Sequential Visit Prediction

We explore the application of advanced neural network techniques in Electronic Health Records (EHR) for disease prediction. Traditional methods using Recurrent Neural Networks (RNNs) have shown promise but are limited in capturing complex clinical relationships within a single patient visit. The paper introduces a novel Multi-Channel Hypergraph Network (MCHN) approach that segments different types of medical codes into multiple channels, applying channel attention to weigh their significance. By constructing a hypergraph from the co-occurrence of medical codes, the model captures high-order interactions, enhancing the predictive power of disease diagnosis. Additionally, the integration of clinical text representations, extracted and encoded using a large language model, further refines the patient's health profile. The proposed MCHN method demonstrates improved performance in sequential visit prediction tasks, offering a more nuanced understanding of patient health trajectories and disease progression.