Graph-Based Diagnostic Prediction Model Based on Global Visit Contexts

In the healthcare field, predicting future diagnoses of patients based on their medical history is a critical task. Electronic health records (EHR) have facilitated the development of many deep-learning models for prediction in this field. However, two issues have not been addressed simultaneously, which can impact the accuracy of the prediction model. The first issue is that rare diseases have little opportunity to be learned during the training process, while the second issue is that short medical visit record sequences pose a challenge for sequential models. To address these challenges, we propose using a graph neural network to encode medical ontology and co-occurrence information into diagnosis representation. We also use a pre-trained vanilla prediction model to obtain similarities between visit contexts and extend visit records by referencing similar visit contexts in the training dataset. Our experiments show that the proposed model performs better than the current state-of-the-art in diagnostic prediction.