Studi Komparatif Random Forest dan Regresi Logistik Elastic Net pada Klasifikasi Status Pulang FKTP BPJS Kesehatan: Kajian F1-Score, PR-AUC, dan Brier Score

Abstract

This study analyzes and compares the performance of Random Forest and Elastic Net Logistic Regression (SAGA solver) for classifying patient discharge status at BPJS Kesehatan Primary Care Facilities (FKTP). The dataset is large-scale, contains exclusively nominal predictors, and exhibits class imbalance (approximately 64.9% majority class). The experimental design employs an 80%/20% train–test split, one-hot encoding for preprocessing, and class balancing on the training data via random undersampling. Hyperparameter tuning is conducted using a staged coarse-to-fine search with local-optimum convergence criteria (improvement threshold ε = 10^(−6) and patience = 10), followed by 10-fold cross-validation for internal evaluation and final assessment on the test set. The primary evaluation metrics are F1-Score, Precision–Recall AUC (PR AUC), and Brier Score. On the test set, both methods achieve identical F1-Scores and nearly identical PR-AUC values. Random Forest and Elastic Net Logistic Regression both attain F1 = 0.996679, with PR-AUC = 0.999931 for Random Forest and 0.999927 for Elastic Net Logistic Regression. A more interpretable difference is observed in probability calibration, where Elastic Net Logistic Regression yields a lower Brier Score (0.002016) compared to Random Forest (0.002706). These results indicate that while both models lie on the same performance plateau in terms of discrimination and ranking, Elastic Net Logistic Regression provides better-calibrated probability estimates.