ADS Capstone Chronicles Revised

23

Figure 8 ROCCurve

XGBoost and Neural Networks also demonstrated strong performance, with high ROC-AUC and PR-AUC scores. While these models effectively handled complex data patterns and class imbalances, their computational complexity and resource demands made them less practical compared to the Bagging Classifier. The Support Vector Machine with the Kernel Trick achieved solid accuracy and was effective in managing complex decision boundaries. However, its performance did not surpass that of the Bagging Classifier in terms of overall effectiveness for this specific application. AdaBoost and Stochastic Gradient Descent, although providing reasonable performance, did not perform as well as the Bagging Classifier. AdaBoost faced challenges in handling the minority class, while SGD’s performance was adequate but not as impactful. The Ridge and Lasso classifier models served as useful benchmarks, providing consistent performance, and establishing a baseline for comparison. Their roles in feature selection and multicollinearity management were beneficial but did not match the advanced capabilities of the Bagging Classifier.

In summary, the Bagging Classifier’s superior performance across multiple metrics makes it the most effective model for detecting FWA in healthcare data. Its ability to manage class imbalance, reduce overfitting, and deliver high accuracy makes it the optimal choice for this project. Future work will focus on leveraging these insights to refine and enhance fraud detection capabilities, potentially integrating strengths from other high-performing models. Overall, the findings indicate that the Bagging Classifier model is the most effective model for detecting FWA within the healthcare system, with SVM providing strong support. Future work will benefit from leveraging these insights to refine and enhance the model further, with a focus on integrating the strengths of the best-performing models to improve fraud detection capabilities. A Flask web application was created to deploy the trained model, enabling users to make predictions through a user-friendly

147