Early identification of neoadjuvant therapy non-response via multimodal immune-imaging biomarkers in breast cancer.

Background: Early identification of breast cancer patients unlikely to benefit from neoadjuvant therapy (NAT) remains a critical unmet need. This study aimed to develop and internally validate a multimodal prediction model for NAT non-response by integrating clinicopathological, tumor microenvironment (TME), longitudinal magnetic resonance imaging (MRI), and systemic inflammatory features. Methods: In this retrospective study, 112 patients with primary breast cancer underwent baseline MRI, a second MRI after two NAT cycles, and definitive surgery. Non-response was defined as Miller-Payne grades 1-3. Candidate predictors were categorized into four domains. After univariate screening, domain-specific multivariable logistic regression was performed, and retained variables entered least absolute shrinkage and selection operator (LASSO) regression to construct a final multimodal model. Internal validation included five-fold cross-validation and 500-iteration bootstrap. Calibration and decision curve analyses were also performed. Results: Thirty-eight patients (33.9%) were non-responders. The individual domain models achieved apparent AUCs of 0.844 (clinical), 0.786 (imaging), 0.828 (TME), and 0.706 (inflammatory). Following LASSO selection, nine features were retained: HER2 status, ER status, Ki-67 index, late enhancement rate after two cycles (LER2), baseline background parenchymal enhancement (BPE), time to peak after two cycles (TTP2), tumor-stroma ratio (TSR), tumor-infiltrating lymphocytes (TILs), and pan-immune-inflammation value after two cycles (PIV2). The multimodal model yielded an apparent AUC of 0.933 (95% CI: 0.890-0.977), with a bootstrap-corrected AUC of 0.855 and a mean five-fold cross-validation AUC of 0.908 ± 0.038. TILs, TSR, PIV2, and Ki-67 were independent predictors. The model demonstrated acceptable calibration after correction for optimism and a net clinical benefit across a range of thresholds. Conclusions: A multimodal prediction model integrating clinicopathological, imaging, tumor microenvironment, and systemic inflammatory features showed potential for early identification of breast cancer patients unlikely to benefit from neoadjuvant therapy. However, given the limited sample size and exploratory single-center design, performance estimates should be interpreted cautiously, and external validation is essential.