Detecting Medicare Fraud: A Machine Learning Approach (#1758)

Authors

Rivas Matta, Rodolfo

Abstract

Medicare fraud detection has become more challenging and urgent over time. It has become urgent because, as some sources suggest, US spending on the Medicare program reached a trillion dollars in 2023, while other reports indicate that the US has lost more than a billion dollars due to Medicare fraud in 2024. At the same time, it has become more challenging because data becomes larger yearly and suffers significant imbalance. Bauder and Khoshgoftaar [1] provide an excellent introductory paper to the problem, explicitly covering the imbalance issue. However, many more problems affect the task of fraud detection, such as data quality, variety of models' configuration, and proper validation of results. As a survey among multiple research studies in this area, this paper has encountered recurring findings and challenges relevant to future work, and it offers suggestions to approach the problem. For example, there are inherent properties in the data collection method used by most studies that provide unrealistic validation results. However, there are sampling methods that prove reliable in getting robust machine learning models regardless of the learners used. The existing studies also indicate the need for a more extensive survey with appropriate comparison methods, as existing works that usually suggest some models perform better (primarily due to tuning hyperparameters) ignore adjusting the competing models with configurations other studies have found important. In addition to other suggestions, this paper highlights the importance of industry collaboration in getting more realistic results and practical models.