This study proposes a novel Dynamic Weighted Selective Ensemble Classifier for robust classification across diverse biomedical datasets. The model integrates bagging with a dual-layer fusion approach–global fusion assigns dynamic weights to base classifiers using a trust evaluation mechanism based on Dempster-Shafer theory and a modified tangent function, while local fusion performs dynamic ensemble selection for context-aware decision refinement. The final prediction is derived through a weighted voting strategy. The efficacy of the proposed model was first validated on 15 publicly available benchmark datasets, where it achieved the best performance in 10 cases, demonstrating strong generalization and adaptability. As a clinical application, the proposed model was evaluated for early response prediction in breast cancer using [ \(^{18}\) F]-FDG PET/CT radiomic features. A total of 81 PET and 66 CT features were extracted per tumor lesion, followed by hybrid feature selection. The proposed model achieved balanced accuracy of 98.72%, 98.38%, 100%, 98.86%, 98.86%, and 100% on the HYPORT PET, HYPORT CT, HYPORT Combined, HYPORT-B PET, HYPORT-B CT, and HYPORT-B Combined datasets, respectively.