Investigating Deceptive Medicare Practices with Official Government Statistics.
In a recent project, researchers delved into the world of Medicare fraud detection using two Logistic Regression models. The goal was to understand feature importances and overall performance in predicting suspicious activities.
The Medicare program, managed by the Centers for Medicare & Medicaid Services (CMS), releases yearly data for Part D, Part B, DMEPOS, and Medicare Advantage (Part C). These datasets are organised on a unique ID known as the National Provider Identifier (NPI).
The average number of unique Medicare Beneficiaries is a significant predictor of fraud, according to the findings. Interestingly, the gender of the physician also played a substantial role, with male providers being identified as the best predictor of fraudulent activities.
To create a comprehensive dataset, features from Part D, Part B, and DMEPOS were combined, joined together on the NPI. This combined dataset model demonstrated a remarkable AUC of 79%, making it the clear winner using this metric.
The random forest model, in particular, placed a lot of emphasis on the median number of 30-Day Fills, Including Refills, the average/Max number of unique Medicare Beneficiaries, and the standard deviation of the number of 30-Day Fills, Including Refills.
Moreover, the model trained on the combined data has a 85% chance of distinguishing between fraudulent and legitimate physicians. The maximum number of Medicare Part D Claims, Including Refills, was also identified as a top feature in predicting fraud.
As more data is collected yearly, the predictive power of these models is expected to increase, offering a promising future in the fight against Medicare fraud.
It's important to note that federal regulation only began allowing data mining like what was described in this project in 2013 for Medicaid data. The Office of The Inspector General releases a list of providers excluded from the Medicare program, known as the List of Excluded Individuals/Entities (LEIE).
For this project, any physician who was active a year prior to their exclusion end year was labelled as a fraud. The eli5 package was used to display feature importances for the model trained on the combined dataset.
In conclusion, this data-driven approach offers a valuable tool in identifying potential fraudulent activities within the Medicare system. By analysing key features such as the average number of unique Medicare Beneficiaries, the gender of the physician, and the number of Medicare Part D Claims, Including Refills, researchers can help safeguard the Medicare program and ensure its continued effectiveness.
