A machine learning framework for cost-efficient resource allocation: evidence from HER2-positive breast cancer diagnosis in low-resource settings

Abstract Accurate identification of HER2-positive breast cancer is essential for treatment planning; however, diagnostic decision-making in resource-constrained healthcare systems is challenged by class imbalance, which can substantially reduce the detection of clinically important minority-class cases. Although numerous imbalance mitigation techniques have been proposed, their evaluation has focused predominantly on predictive performance, with limited attention to their implications for operational decision support and healthcare resource allocation. This study presents a machine learning framework that integrates predictive evaluation, statistical validation, and operational interpretation to support evidence-based selection of imbalance mitigation strategies for HER2-positive breast cancer diagnosis. A subset of 1400 complete patient records from the METABRIC breast cancer dataset was analyzed using nine clinical predictors and HER2 status as the target variable. Decision Tree, Support Vector Machine (SVM), and XGBoost classifiers were evaluated under four imbalance mitigation strategies: Baseline, Synthetic Minority Oversampling Technique (SMOTE), Cost-Sensitive Learning (CSL), and a combined SMOTE + CSL approach. Model performance was assessed using repeated stratified hold-out validation (30 repetitions) and evaluated using Accuracy, Precision, Recall, F1-score, and ROC-AUC. Statistical significance was assessed using predefined pairwise Wilcoxon signed-rank tests. The results demonstrated that the effectiveness of imbalance mitigation was strongly classifier-dependent. SMOTE and Cost-Sensitive Learning generally improved HER2-positive detection, whereas no single strategy consistently achieved superior performance across all classifiers and evaluation metrics. SVM exhibited the greatest responsiveness to imbalance mitigation, while Decision Tree remained comparatively stable and XGBoost showed more moderate improvements. Importantly, the combined SMOTE + CSL strategy produced perfect recall for SVM but substantially reduced precision and overall accuracy, demonstrating that uncalibrated combinations of imbalance mitigation techniques may overcompensate the learning objective and produce degenerate prediction behavior despite apparently favorable sensitivity. Statistical analysis identified statistically significant differences for predefined pairwise comparisons across repeated experiments, supporting the robustness of the comparative evaluation. By integrating predictive performance, statistical robustness, and operational interpretation within a unified evaluation framework, this study extends conventional assessments of class imbalance mitigation beyond predictive accuracy alone. The proposed framework supports evidence-based diagnostic prioritization and provides practical guidance for deploying machine learning models in resource-constrained healthcare settings where balancing minority-class detection with efficient use of diagnostic resources is essential.