Hierarchical Graph Neural Network for Patient Treatment Preference Prediction with External Knowledge

The healthcare industry has a wealth of data that can be used by researchers and medical professionals to infer a patient’s condition and intention to receive treatment using machine learning models. However, this line of research generally suffers from some limitations: (1) struggling to leverage structural interactions among patients; (2) attending to learn patient representations from electronic medical records (EMRs) but rarely considering supplementary contexts; and (3) overlooking EMR data imbalance issue. To address these limitations, in this paper, we propose a hierarchical graph neural network for patient treatment preference prediction. Doctors’ information and their viewing activities are first integrated as external knowledge with EMRs to construct the hierarchical graph, where a dual message passing paradigm is then devised to perform intra- and inter-subgraph aggregation to enrich patient representations and advance label propagation. To mitigate patient data imbalance issue, a community detection method is further designed to better prediction. Our experimental results demonstrate the state-of-the-art performance on patient treatment preference prediction.