Heterogeneous Patient Graph Embedding in Readmission Prediction

The number of youth seeking mental health services has been increasing in the past decade. Accurate prediction of hospital readmission is a contributing factor in addressing youth mental health problem and healthcare service utilization. Medical records are an important source of information for readmission prediction, however utilizing such records in the context of mental care, requires overcoming two impeding challenges: the diversity of service utilization (e.g., using psychiatric vs. overdose vs. trauma clinics all for the same underlying mental health reason) and the heterogeneity of data associated with each service. Graph Neural Network (GNN) is shown to be promising in performing classification or regression tasks when input data bears a complex structure. In this research, we propose using graph embedding to first generate patient graph that captures episodic emergency department visits and the complex service utilizations of the patient. We then use GNN for readmission prediction. For embedding and training purposes, we utilize more than 4,000 unique mental health patients data over 19 years. To evaluate our approach we systematically compare a variety of of GNN models with four RNN models, namely LSTM, Bi-LSTM, GRU, and Bi-GRU. Our experimental evaluation demonstrates that encoding the complex interrelationship between features of a patient using graph embedding and GNN improves the performance of the predictive model compared to RNN counterparts.