INFER: A Multi-Agent Framework for Detecting Fraud, Waste, and Abuse in Medicare Claims

Srinidhi Moodalagiri, Amber Nigam, Arpan Saxena, Shen Wang, Dan Roth

Keywords: Healthcare fraud detection, Multi-agent systems, Medicare claims analysis, Agentic AI, Program integrity, Anomaly detection

Abstract: Medicare Fraud, Waste, and Abuse (FWA) costs the U.S. healthcare system between \\$100B and \\$300B annually, yet current detection approaches recover only a fraction of these losses. We identify three structural gaps in existing methods that limit their effectiveness and present INFER, a multi-agent platform that addresses each gap through purpose-built architectural components. Applied to the CMS 5% Limited Data Set (LDS) spanning 3.6M DME claims, 254K hospice claims, and 42.8M Part B claim lines (2022–2024) as part of the CMS Chili Cook-Off Challenge, INFER achieved a precision of 0.98 on clinician-reviewed high-confidence cases, an approximately 55% reduction in false positives attributable to the multi-agent design, and 9 fully executed FWA detection pipelines. We detail the FWA taxonomy co-designed with clinical experts, present ablation analysis quantifying each agent’s contribution, and provide worked examples demonstrating how the multi-agent design detects patterns invisible to prior approaches.

Submission Number: 11