A robust and interpretable ensemble machine learning model for predicting healthcare insurance fraud

Area of Science:

• Computer Science
• Healthcare Management
• Data Science

Background:

• Healthcare insurance fraud causes billions in global losses annually.
• Accurate fraud detection is crucial for financial sustainability in healthcare.
• Existing methods often lack interpretability and optimal feature selection.

Purpose of the Study:

• To improve the accuracy of healthcare insurance fraud detection.
• To enhance the interpretability of machine learning models used for fraud detection.
• To identify optimal feature subsets for efficient model performance.

Main Methods:

• Data preprocessing and feature selection using embedded and permutation methods.
• Application of ensemble techniques (Voting, Weighted, Stacking) for model aggregation.
• Feature interpretation using Partial Dependence Plots (PDP), SHAP, and LIME.

Main Results:

• Identification of minimal feature sets achieving high fraud detection performance.
• Demonstration of improved accuracy through ensemble machine learning methods.
• Successful interpretation of feature importance and impact on predictions.

Conclusions:

• Machine learning, particularly ensemble methods, offers a powerful approach to combat healthcare insurance fraud.
• Feature selection and interpretation are key to developing robust and understandable fraud detection systems.
• This research provides a framework for more accurate and interpretable fraud detection in healthcare.