Analysis of ML Models & Data Balancing Techniques for Medicare Fraud Using PySpark

Medicare fraud is a persistent and costly issue, contributing to billions of dollars in annual losses in healthcare systems worldwide. With the increasing complexity of healthcare data and fraudulent schemes, traditional detection methods struggle to keep up, making Big Data analytics and machine learning essential for effective fraud detection. This paper proposes a robust methodology for detecting Medicare fraud using a PySpark framework, leveraging datasets from the CMS, LEIE, FDA. The approach integrates these datasets to create a comprehensive profile for each claim and applies feature selection to identify indicators of fraudulent behavior. The methodology employs data balancing techniques, such as RUS and SMOTE to address class imbalances in fraud data. Various machine learning models, including Gradient Boosted Trees, Random Forest, and Logistic Regression, are trained and evaluated on their accuracy, precision, recall, and specificity scores. The experimental results show that the proposed models effectively identify fraudulent claims, with the model trained on data balanced at a 1:5 ratio of fraud to non-fraud cases achieving the highest specificity of 73%. This work not only improves Medicare fraud detection but also offers a scalable approach that can be applied to large-scale healthcare schemes globally, such as India's Ayushman Bharat program. The results underscore the potential of Big Data solutions in safeguarding healthcare resources by mitigating fraud.