Data mining supported causality analysis of outpatient medical treatment reimbursements

Many insurance claims are fraudulent. One part of potential fraudulent or abusive claims in the domain of Swiss accident insurances are non-causal outpatient medical treatment reimbursements. The identification of these invoices is still an exhausting manual task which is only partly supported by IT-systems. This thesis shows a general procedure based on existing data mining techniques to support the causality analysis of outpatient invoices. This thesis is based on a dataset obtained from the largest Swiss accident insurance company (Suva). The analysis showed that a labeled non-causal invoice is a very rare event with an abusive/fraudulent rate around one-fourth of a percent. This significant difference in class prior probabilities is known as class imbalance problem. The designed general procedure shows how to prepare and transform the available invoice and claim notification information into a classification problem as well as how to handle the class imbalance problem. The goal of this study was to find an existing data mining algorithm that has a positive effect on both effectiveness (which invoice is deflected for manual validation) and efficiency (why is invoice deflected). However, many classification algorithms lacks especially in the interpretability of the generated classification model. This thesis shows that a data mining technique based on classification rules derived from frequent patterns achieves the best results regarding interpretability and prediction accuracy....