Can Machine Learning Target Health Care Fraud? Evidence from Medicare Hospitalizations

Area of Science:

• Health economics
• Medical informatics
• Machine learning

Background:

• US healthcare spending exceeds $4 trillion annually, with significant concerns about hospital overbilling and fraud.
• Private providers and insurers dominate healthcare, creating incentives for inflated claims.
• Existing fraud detection methods often lack efficiency and transparency.

Purpose of the Study:

• To develop novel machine learning tools for identifying hospitals engaged in overbilling and fraud.
• To guide investigations and audits within public and private health insurance systems.
• To provide explainable insights into the diagnosis, procedure, and billing codes associated with suspicious claims.

Main Methods:

• Utilized large-scale Medicare inpatient hospitalization claims data.
• Developed a fully unsupervised machine learning approach for fraud detection.
• Incorporated explainability features to identify contributing billing codes.
• Validated findings using Department of Justice data on anti-fraud lawsuits and case studies.

Main Results:

• Identified patterns consistent with fraud in inpatient hospitalizations.
• The unsupervised method achieved a nearly 5-fold lift over random hospital targeting.
• The approach provides interpretable results, highlighting specific codes linked to suspiciousness.
• Post-analysis revealed associations between certain hospital characteristics and suspiciousness.

Conclusions:

• Novel machine learning tools can effectively detect hospital overbilling and fraud in Medicare claims data.
• The unsupervised and explainable nature of the method enhances its utility for auditing and investigation.
• Findings support the integration of advanced analytics for improving healthcare payment integrity.