Applications of Machine Learning to the Identification of Anomalous ER Claims

Abstract: Improper health insurance payments resulting from fraud and upcoding result in tens of billions of dollars in excess health care costs annually in the United States, motivating machine learning researchers to build anomaly detection models for health insurance claims. This article describes two such strategies specifically for ER claims. The first is an upcoding model based on severity code distributions, stratified by hierarchical diagnosis code clusters. A statistically significant difference in mean upcoding anomaly scores is observed between free-standing ERs and acute care hospitals, with free-standing ERs being more anomalous. The second model is a random forest that minimizes improper payments by optimally sorting ER claims within review queues. Depending on the percentage of claims reviewed, the random forest saved 12% to 40% above a baseline approach that prioritized claims by billed amount.