Predicting risk of Rejection in Non-Submitted claims

Abstract

A 2011 study by the U.S. Government Accountability Office found that claim denial rates vary significantly among states and health insurers. Of the small number of states tracking such information, denials ranged between 11 percent and 24 percent of claims. The following are results from the National Health Insurer Report Card (NHIRC) years 2008-2013 that address denials. Percentages of claim lines denied: What percentage of claim lines submitted are denied by the payer for reasons other than a claim edit? A denial is defined as: allowed amount equal to the billed charge and the payment equals $0. Hence if a system is in place to predict the risk of rejection much before the claim is actually submitted, the percentage of rejection of claims for a provider can be reduced and its revenue improved. The problem is modeled as supervised classification problem, training data consisted of all the labelled claims and model accuracy is checked on test data set. Overall data set was spread among multiple tables, proper understanding of features and data preprocessing lead us to combine the tables in sensible manner over which model can be built. Multiple models are built with proper tuning, and their performances are compared. We found Decision Tree and Logistic Regression as top models which were giving around 90% accuracy.