ENSEMBLE MODEL FOR FRAUD DETECTION IN HEALTH INSURANCE CLAIMS BASED ON TRANSACTION DATASET AND MEMBER BEHAVIOR USING LOF, IFOREST, AND Z-SCORE

• Main Author: Alit Cahya, Neo
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/87576
• Institution: Institut Teknologi Bandung

Fraud in health insurance claims is a significant challenge that affects the financial stability and sustainability of insurance companies' services. The main contribution of this study is to design a fraud detection framework in health insurance claims based on unsupervised learning with an ensemble approach that combines the Local Outlier Factor (LOF), Isolation Forest (iForest), and Z-Score algorithms. The dataset used consists of insurance claim transaction data and member behavior, which are processed to detect anomalies through a combination of ensemble methods such as All Vote and Any Vote. The evaluation matrices used are Silhouette Score, Davies-Bouldin Index, and Calinski-Harabasz Index to assess the quality of the clustering. The results of the study show that the ensemble method can improve model performance, namely the All Vote ensemble which has the best performance with a Silhouette Score value of 0,94, Davies-Bouldin Index of 0,66, and Calinski-Harabasz Index of 115.160,38. The results of the evaluation of this framework with BestModel AnyVote can detect up to 6 cases of fraud, while conventional methods are unable to identify any. This framework is expected to be an innovative solution in detecting fraud earlier, increasing the operational efficiency of insurance companies, and minimizing losses due to fraud.