Predicting the Next Diseases Using Graph Neural Networks on Administrative Medical Datasets

This paper aims to address the problem of next disease prediction using advanced graph-based methods on administrative medical datasets. The objective of our study is to predict the next possible disease based on a patient’s past disease records. Traditional statistical methods have been used to measure disease associations in administrative medical datasets, but graph-based methods have emerged as a promising approach for analyzing such datasets. The proposed method uses Gated Graph Neural Networks (GGNN) to analyze both past and recent medical conditions, to uncover latent patterns and connections within a patient’s medical history. By utilizing the session graph embeddings, past diseases can be identified, while global graph embeddings can be used to predict future diseases that the patient is likely to develop. A soft-attention mechanism is also employed to combine both global and local information, resulting in accurate predictions of future related diseases based on the patient’s medical history. Our proposed method demonstrates superior performance compared to several baseline approaches in predicting the next diseases, highlighting its effectiveness in modeling the relationship between diseases using a graph-based approach.