Investigating US Healthcare Referral System with Graph Neural Networks

To understand the referral pattern in the healthcare system, I construct the patient-sharing network and explore its predictive power of network structure on future referrals. Using the data from the Centers for Medicare & Medicaid Services (CMS), which contains a 5% random sample of Medicare beneficiaries in 2008 and their insurance claims from 2008 to 2010, I can analyze how patients' cost of treatment varies according to different factors. While constructing a large-scale network based on the data, I employ different graph neural network algorithms, including Node2vec, Attri2vec, and graphSAGE, to extract link predictions and node embeddings from the network structure. Moreover, I compare the three algorithms' accuracy, concluding that Attri2vec has the highest accuracy when applying the model to our test dataset. Last, I show that medical referrals can be automated for efficient mechanism design for cost-saving and efficiency improvement as well as promoting diversity and inclusions. In conclusion, the node embeddings of prediction power show how useful they are for describing the referral pattern through time. These embeddings successfully identify potential connections among physicians and patients that open up new avenues for the referral system.