An Investigative Analysis of PSO‐Optimized Multiclassifier Ensemble for Reliable Breast Cancer Diagnosis
Narambunathan Arunachalam Natraj, Venkatasubramanian K., Pankaj Pathak, Sandeep Prabhu, S. Gopinath · Applied Computational Intelligence and Soft Computing · 2025
The high mortality rate of breast cancer necessitates improved early detection methods to enhance survival rates. This study addressed the limitations of traditional, often subjective diagnostic techniques by developing a machine‐learning approach for more accurate and efficient breast cancer detection. Using the Wisconsin Breast Cancer dataset, the proposed method investigated the efficacy of ensemble learning techniques in classifying cases as benign or malignant. The proposed methodology employed a multistage machine learning pipeline incorporating six classifiers: support vector machine, decision tree (DT), random forest, logistic regression, AdaBoost, and XGBoost, combined through a voting classifier optimized by particle swarm optimization (PSO). The research evaluated performance using accuracy, precision, recall, F1‐score, specificity, area under the curve (AUC), log loss, false discovery rate (FDR), and false omission rate (FOR). Results showed varying performance across individual classifiers, with XGBoost achieving the highest testing accuracy (0.9825), followed by logistic regression (0.9766). The PSO‐optimized voting classifier matched logistic regression’s accuracy (0.9766) and demonstrated balanced performance across metrics (precision, recall, F1‐score: 0.9683; AUC: 0.9962). While not surpassing XGBoost, the ensemble approach offered more consistent performance than most individual models, effectively mitigating the weaknesses of classifiers like DT (accuracy: 0.9415) and AdaBoost (accuracy: 0.9298). Ensemble methods enhanced overall robustness in breast cancer detection, balancing high accuracy with consistent performance across various metrics. The study revealed a trend toward improved performance through model combination and optimization techniques. However, the marginal improvement over the best individual classifier (XGBoost) suggested room for further optimization. The research concluded that the PSO‐optimized ensemble approach provided a promising solution for the detection of breast cancer, potentially supporting more reliable clinical diagnostics. Future research should refine ensemble weights, incorporate relevant features, and validate the model on varied datasets to increase performance and generalizability.