Exploring Graph Neural Network in Administrative Medical Dataset

Abstract

Administrative medical dataset contains medical records of patients. Using administrative medical dataset can extract disease association to help finding comorbidity. Previous methods only use statistics on administrative medical dataset such as computing probabilities of disease occurrence and are limited by the capability of statistics. To enhance hidden information usage of administrative medical dataset, we propose two different methods based on graph neural networks to exploit hidden information in administrative medical dataset. One is using graph neural networks to generate disease embeddings and pass through kNN algorithm to find similar diseases for suggestion to physicians. The other one is that we formulate sequence prediction problem and use gated graph neural network to model every disease sequence by forming session graphs. Different from previous methods for sequence prediction that only consider current sequence, we also consider all sequences in dataset at the same time. Besides, we use position-aware soft-attention mechanism to aggregate disease embeddings to session embeddings and predict the next disease of a patient. We conduct extensive experiments on two methods and show its ability to outperform several baselines.