Baseline and Early Treatment MRI Model for Predicting Complete Pathologic Response to Neoadjuvant Chemoimmunotherapy in Patients With Triple-Negative Breast Cancer

Background: Triple-negative breast cancer (TNBC) is a highly aggressive breast cancer subtype lacking targeted therapies. Neoadjuvant chemoimmunotherapy (NACI) improves pathologic complete response (pCR) rates, although patient selection is challenging. Objective: To develop and test a model incorporating baseline and early-treatment MRI features, including dynamic contrast enhancement (DCE) features, for predicting pCR in patients with TNBC undergoing NACI. Methods: This retrospective study included patients with TNBC undergoing NACI and who underwent breast MRI including DCE before treatment and after the first treatment cycle (i.e., early NACI), including a single-center training set of 90 women (mean age, 49 years; January 2018 to September 2024) and an external test set of 29 women (mean age, 46 years; date range unavailable) from publicly available trial data. Two radiologists evaluated MRI features including percentage enhancement (PE) reduction, representing semiquantitative assessment of relative expansion of intralesional nonenhancing components after early NACI. A model for predicting pCR on definitive surgery after NACI completion was constructed in the training set using independent predictors from multivariable logistic regression analysis and was evaluated in the external test set. Shapley additive explanations (SHAP) analysis was used to identify features' contributions to model predictions in the training set Results: Independent predictors of pCR in the training set were tumor unifocality (OR=7.2, p=.001) on pretreatment MRI and early tumor shrinkage (ETS) ≥37% (OR=9.7, p<.001) and PE reduction (OR=9.7, p<.001) on early-NACI MRI. A model incorporating these parameters achieved in the external test set AUC of 0.88, sensitivity of 74%, and specificity of 90% for predicting pCR. In the external test set, calibration curves showed strong concordance between model-predicted and observed pCR outcomes, and the Hosmer-Lemeshow test showed satisfactory model fit (p=.67). In SHAP analysis, global importance for model predictions was highest for PE reduction (mean absolute SHAP value, 0.42), followed by ETS (0.32) and unifocality (0.21). Conclusion: A clinically practical model was created for early pCR prediction in patients undergoing NACI for TNBC. Clinical Impact: This MRI-based predictive model could facilitate timely tailoring of clinical regimens after immunotherapy initiation by informing optimal de-escalation strategies for responders while prompting therapeutic adaptations for nonresponders.