Catch me if you can - How to use AI to fight Fraud, Waste and Abuse (FWA)?
summary
Workflow combines Medicare Provider Utilization and Payment Data (2012–2017) with the LEIE excluded individuals and entities list via NPI to label ~54 million claims of which ~4.7k become fraud. Models use random undersampling and a TensorFlow boosted tree classifier with 50/50 class balance, yielding ~76% accuracy, ~70% precision, 85% recall and 0.84 AUC. Also experiments with GANs and one-class adversarial nets that need little or no fraud examples and detect more emerging abuse. Source code and data links are open.
AI-written summary · about
chapters (12)
- 0:00Start
- 7:07Fraud, Waste, Abuse and Error
- 15:29List of Excluded Individuals and Entities (LETE)
- 18:53Final data set structure
- 19:55Confusion Matrix
- 23:17Random Under Sampling (RUS)
- 25:28How about other models?
- 25:44Model performance comparison
- 27:27What define a Fraud Detection System gooc- or bad?
- 32:06Where is the Outlier?
- 32:41Unsupervised Clustering
- 35:33Nearest Neighbor Approach
related talks
16mGenAI in Healthcare Claims: Processing and Detecting Fraud
17mAI-Ready Healthcare Data: Revolutionizing Fraud Detection
35mGenerative AI for Fraud Detection & Claims Analysis
16mGenerative AI for Claims and Fraud
38mFinding Fraud Indicators in Medical Insurance Claims
35mAI Powered Fraud Detection for Payments & Transactions
16mAI-Driven Claims Adjudication: Healthcare Efficiency & Fraud Prevention
08mAI-Powered Fraud Detection in Financial System