Scholarlink

Fraud in medical insurance claims accounts for approximately 5–10% of total claim costs, posing significant financial and reputational risks for insurers. Traditional detection methods, such as manual audits, are inefficient and prone to errors, necessitating the adoption of machine learning-based approaches. This study proposes a hybrid ensemble model for fraud detection, integrating Support Vector Machines, Random Forest, and K-Nearest Neighbors, with a Gradient Boosting meta-model for optimized predictions. The proposed method consists of four main phases: preprocessing (scaling, encoding, and class balancing using SMOTE), modeling (training individual classifiers), optimization (meta-model integration), and evaluation using Accuracy, F1-score, and RMSE. The model was validated on a real-world medical insurance claims dataset, demonstrating a 2.93% increase in accuracy, a 10.37% reduction in RMSE, and a 14.41% improvement in MAE compared to the best individual model. The main contribution is a scalable and interpretable hybrid system that enhances fraud detection in insurance. Future research will focus on expanding feature sets, incorporating external data sources, and applying interpretability techniques such as SHAP to improve transparency and adoption in real-world applications.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Novel Fraud Detection Method in Medical Insurance Claims Using Ensemble Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • David Galarza,
  • Rodolfo Bojorque,
  • Remigio Hurtado

摘要

Fraud in medical insurance claims accounts for approximately 5–10% of total claim costs, posing significant financial and reputational risks for insurers. Traditional detection methods, such as manual audits, are inefficient and prone to errors, necessitating the adoption of machine learning-based approaches. This study proposes a hybrid ensemble model for fraud detection, integrating Support Vector Machines, Random Forest, and K-Nearest Neighbors, with a Gradient Boosting meta-model for optimized predictions. The proposed method consists of four main phases: preprocessing (scaling, encoding, and class balancing using SMOTE), modeling (training individual classifiers), optimization (meta-model integration), and evaluation using Accuracy, F1-score, and RMSE. The model was validated on a real-world medical insurance claims dataset, demonstrating a 2.93% increase in accuracy, a 10.37% reduction in RMSE, and a 14.41% improvement in MAE compared to the best individual model. The main contribution is a scalable and interpretable hybrid system that enhances fraud detection in insurance. Future research will focus on expanding feature sets, incorporating external data sources, and applying interpretability techniques such as SHAP to improve transparency and adoption in real-world applications.