A medical treatment based scoring model to detect abusive institutions.

Abstract

Medical abuse refers to a type of abnormal medical practice which is not in compliance with qualitative or ethical standards, such as excessive prescription or overbilling of medical services. Detection of such medical abuses is crucial, especially for the patients and insurance providers, because they become subject to the extra payments incurred. As a result, insurance providers hire medical experts in order to review claims manually, yet through examination is almost impossible due to the volume of the claims filed. A typical approach is to select institutions on suspicion of abusive practices and to manually review all claims involving suspect institutions. In this light, several studies have developed models designed to extract institution-level variables. However, since these variables are at an institution-level, the model cannot account for different types of abuse practiced by individual institutions, hence degrading the accuracy of the prediction model. At the same time, these variables contain information too simple to construct an effective scoring model. In this study, we propose a model that scores the degree of abuse practiced by institutions at the treatment-level, which is the lowest level of data that can be obtained from a medical claim. Our model is the first to use such fine-grained information to construct a model for scoring the abuse by medical institutions. The proposed model consists of two stages: Training a deep neural network with embedding layers for categorical variables, and scoring the abuse degree for each treatment with the model. Then, we aggregate the resulting abuse score of each treatment and the claim amount associated with each treatment by an institution which we define as the abuse score of the institution. We test our model using real-world claim data submitted to the Health Insurance Review and Assessment (HIRA) in 2016. We also compare the performance of the proposed model against the scoring model HIRA has been using, which computes the abuse score of an institution by using institution-level variables as proposed in past literature. Experiment results show that the proposed model represents the degree of medical abuse better. In addition, the results suggest that the reviewers may examine through the claims by at most 6.1 times more efficiently than when using the scoring model with institution-level variables.