Multi-Parametric Ultrasound Radiomic Kinetics with Machine Learning Ensemble for Early Prediction of Pathologic Response to Neoadjuvant Chemotherapy in Breast Cancer

Background/Objectives: Early identification of breast cancer patients unlikely to benefit from neoadjuvant chemotherapy (NAC) remains a pressing clinical problem because ineffective therapy delays definitive surgery and exposes patients to unnecessary toxicity. Quantitative ultrasound (QUS) and shear wave elastography (SWE) probe complementary tissue properties—scatterer microstructure and mechanical stiffness—that may change before macroscopic tumor shrinkage. This study aimed to evaluate whether multi-parametric ultrasound (mpUS) radiomic kinetics, analyzed with a machine learning ensemble and interpreted with SHAP, could predict pathologic response to NAC. Methods: A prospective observational cohort enrolled 135 women with biopsy-proven stage II–III breast cancer treated with NAC within the multidisciplinary breast pathway shared between Vasile Goldis Western University of Arad and Victor Babes University of Medicine and Pharmacy Timisoara (Pius Brinzeu County Emergency Hospital). All patients underwent standardized QUS and SWE acquisitions at baseline, week 1, and week 3. Response was defined pathologically at surgery as residual cancer burden (RCB) class 0/I versus II/III. Group comparisons used Welch’s t-test, Mann–Whitney U, chi-square, and Fisher’s exact tests; correlations used Spearman’s rho. A stacked machine learning ensemble (four base learners—XGBoost, random forest, support vector machine, and L2-penalized logistic regression—combined by a separate second-stage logistic meta-learner) was trained with nested 10-fold cross-validation, bootstrap stability assessment, SHAP-based interpretability, and decision curve analysis. Results: Sixty patients (44.4%) were responders and 75 (55.6%) were non-responders. Responders showed greater week 3 increases in mid-band fit (3.4 ± 0.9 vs. 1.2 ± 0.8 dB, p < 0.001), entropy (0.7 ± 0.2 vs. 0.2 ± 0.2, p < 0.001), and more pronounced SWE mean stiffness reduction (−44.1 ± 9.7 vs. −12.1 ± 8.6 kPa, p < 0.001). The stacked ensemble integrating clinical, QUS, and SWE kinetic features reached an AUC of 0.93 (95% CI 0.88–0.97) versus 0.71 for the clinical-only model (all reported performance figures represent internal cross-validation only). SHAP analysis identified Δ MBF and Δ entropy at week 3 as the dominant features, with high bootstrap stability. Conclusions: Multi-parametric ultrasound radiomic kinetics integrated through a machine learning ensemble may provide an interpretable early-response biomarker for NAC in breast cancer, pending external validation.