An Ensemble Approach for Inconsistency Detection in Medical Bills: A Case Study
Gianlucca Lodron Zuin, Lucas Parreiras, Luiz Ledo Mota Melo, Gabriel Barros, Humberto Lomeu, Batielhe Melo, Wesley Marini, Debora Lott, Mateus De Souza · 2023
Auditing medical bills is a complex task for large health insurance companies, such as Unimed-BH, which serves over 1.5 million patients. To overcome this challenge, our study proposes an ensemble-based tool that ranks medical bills based on their likelihood of being inconsistent. Our ensemble comprises multiple machine learning algorithms and employs a Multi-Armed Bandit Algorithm to optimally combine its constituents. Using data collected from November 2022 to March 2023, we created a dataset with 81,173 hospital operations (with up to 5,972 items) and 508,304 outpatient operations (with up to 2,402 items). The proposed method proved effective in detecting inconsistencies in both scenarios. By ranking all bills and flagging the top of the ranking as potentially inconsistent, we achieved a recall of. 852 at the 5% threshold. Our proposed solution was deployed in the real world, and over a five-month period, it recovered R$1, 571,146 (USD$349, 610) which would have gone unaudited. Our findings suggest that the proposed approach is effective and can help free up expert auditors to handle more complex tasks, increases efficiency, and reduce costs while detecting inconsistencies in medical bills.