Penggunaan Algoritma Stacking Classifier Pada Sistem Deteksi Risiko Kardiovaskular

Abstract

Cardiovascular disease is a leading cause of global death. However, the complexity of medical data often makes conventional models fail to capture hidden patterns, resulting in suboptimal predictive performance. This study evaluates the effectiveness of a hybrid model that integrates K-Modes Clustering with the Stacking Classifier algorithm and tests whether the model's complexity can provide significant performance improvements compared to a single model. The methodology involves data preprocessing including outlier handling, clinical feature engineering, and cluster feature extraction using K-Modes (K=2). The Stacking Classifier architecture is built using five optimized heterogeneous base-learners (CatBoost, Decision Tree, MLP, SVC, Logistic Regression) and XGBoost as a meta-learner, validated through Stratified 5-Fold Cross-Validation. The results showed that although K-Modes effectively mapped clinically valid risk categories, the Stacking Classifier model (87.99% accuracy and 95.89% ROC-AUC) was not able to surpass the performance of the best single model, namely CatBoost (88.03% accuracy and 95.90% ROC-AUC). The most significant finding lies in the computational time efficiency, where the Stacking Classifier algorithm required 560 times longer computational time (7587.7686 seconds) than CatBoost (13.4635 seconds) without providing a commensurate performance improvement. This indicates that Boosting-based algorithms are able to capture complex patterns without requiring additional ensemble layers, so that an optimized single model is more recommended for real-world implementations by providing the best balance between prediction accuracy and computational time efficiency.