Application of Stacking Ensemble Learning for Classifying Patient Revisit Patterns in Primary Healthcare Facilities

Abstract

The increasing volume of patient visits in primary healthcare facilities (such as community health centers and clinics) demands efficient resource management and scheduling systems. The inability to predict patient arrival patterns often leads to long queues and suboptimal allocation of medical personnel. To overcome these issues, this study proposes the development of an intelligent framework based on Stacking Ensemble Learning to accurately classify and predict patient revisit frequency patterns. The proposed architecture integrates three heterogeneous base classification models (base learners), namely Logistic Regression (LR), Support Vector Machine (SVM) with non-linear Radial Basis Function (RBF) kernel projection, and Random Forest (RF). The probability predictions from these three base models are then combined into a secondary Logistic Regression based meta-learner to optimize the final decision and reduce the predictive bias of single models. Experiments were conducted using a dataset sourced from actual primary healthcare medical records, totaling 499 data entries. Given the highly imbalanced distribution of demographic data, the preprocessing stages were comprehensively designed, encompassing numerical scale feature standardization and the application of the SMOTE (Synthetic Minority Oversampling Technique) class balancing algorithm. Model testing was performed using a 5-fold stratified cross-validation scheme to ensure the robustness of the model's generalization. Comprehensive evaluation results proved that the Stacking Ensemble architecture achieved peak performance with an Accuracy of 88.4%, Precision of 85%, Recall of 90%, and an F1-Score of 0.88. Furthermore, curve analysis revealed an Area Under the Receiver Operating Characteristic (AUC-ROC) score of 0.961 and a Precision-Recall AUC of 0.95, consistently outperforming the capabilities of each individual baseline model. These findings confirm that combining hybrid machine learning models can produce highly stable and reliable predictions while minimizing false negative errors. This architecture holds massive potential for practical implementation as the foundation of a decision support system in managing operational services within primary healthcare facilities.