This study demonstrates that the performance of Medicare fraud detection models can be improved by applying graph neural networks to graph-structured datasets. Tens of billions of dollars are lost to Medicare fraud yearly. Therefore, insurance companies constantly attempt to enhance fraud detection models to alleviate economic losses and reputational damage. While traditional rule-based or machine learning models have been extensively studied for Medicare fraud detection, graph learning has been used in very few studies. Thus, a Medicare fraud detection model that applies the GraphSAGE algorithm, a graph neural network, to graph-structured datasets generated from open-source data, including the information of Medicare providers, physicians, and beneficiaries, is presented in this paper. We developed a heterogeneous graph by setting the Medicare providers and beneficiaries as nodes. Consequently, the GraphSAGE model outperformed the baseline model in terms of precision, recall, F1score, and area under the receiver operating characteristics curve by 0.01, 0.35, 0.30, and 0.18, respectively. This result indicates that the graph relationship between the provider-beneficiary and provider-physician conveys important information that can be exploited for Medicare fraud detection.