Fraud in healthcare claims represents a critical challenge that undermines the efficiency and sustainability of Indonesia's National Health Insurance (JKN) system. This study contributes a large-scale comparative evaluation of five machine learning algorithms for classifying potential fraud in BPJS Kesehatan claims, namely Artificial Neural Network (ANN), Support Vector Machine (SVM), Random Forest (RF), XGBoost + SMOTE, and Logistic Regression (LR). A novelty of this study lies in applying the SMOTE technique in conjunction with XGBoost to address class imbalance in fraud datasets. The dataset consists of over 200,000 claim entries, which have undergone data cleaning, normalization, and feature selection. Performance was assessed using precision, recall on fraud class (positive), f1-score, accuracy, and confusion matrix visualizations to capture classification error distribution. Results demonstrate that ANN and XGBoost + SMOTE are superior in detecting fraudulent claims with high recall, while SVM achieves the most balanced performance in terms of precision and sensitivity. Random Forest and Logistic Regression serve as moderate baselines but are less effective in identifying complex fraud patterns. This study contributes to the development of a more adaptive and efficient fraud detection system based on machine learning, with practical implications for strengthening the automatic verification system used by BPJS Kesehatan.