Abstract No. 
2022 San Antonio Breast Cancer Symposium
8-11 Dec

MRI models by response predictive subtype for predicting pathologic complete response

Li W, Onishi N, Wolf DM, Newitt DC, Yau C, Wilmes LJ, Gibbs JE, Price ER, Joe BN, Kornak J, LeStage B, I--SPY2 Imaging Working Group, I-SPY2 Consortium, Esserman LJ, van 't Veer L, Hylton NM

Background: MRI predictive modeling is used in the I-SPY 2 neoadjuvant clinical trial as a key component of the pre-RCB (Predicted Residual Cancer Burden) clinical workflow for re-directing “good responders” to skip AC (anthracycline) and proceed to surgery early. The current MRI model is hormone receptor (HR)- and human epidermal growth factor receptor 2 (HER2)-specific, and was trained retrospectively using data from 990 patients in I-SPY 2. Recently, new breast cancer subtypes based on gene expression and pathologic response were proposed by Wolf et al [1]. Their study predicted that drug allocation by the new response-predictive subtype (RPS) would lead to a higher pathologic complete response (pCR) rate than allocation based on HR/HER2 subtypes. In this project, we evaluated the MRI model optimized by RPS and compared it with the HR/HER2 optimized model.

Methods: A total of 990 patients enrolled in I-SPY 2 and randomized to one of 9 drug arms or control were evaluated in this analysis. Functional tumor volume (FTV) was calculated from dynamic-contrast enhanced MRI [2] performed pretreatment (T0), after 3 weeks of treatment (T1), and between sequential drug regimens (T2). pCR was assessed at surgery after treatment was completed. HR/HER2 subtype was defined by HR and HER2 +/-, which resulted in four subtypes: HR+/HER2-, HR+/HER2+, HR-/HER2+, and HR-/HER2- (triple negative). RPS subtype was defined by immune, DNA repair deficiency (DRD), HER2, and BluePrint (BP) subtype (Agendia) biomarkers to define five subtypes: HER2-/Immune-/DRD-, HER2-/Immune+, HER2-/Immune-/DRD+, HER2+/BP-HER2_or_Basal, and HER2+/BP-Luminal. A logistic regression model using at least 1 FTV variable (value at T0, percent change at T1 or T2) was analyzed for predicting pCR. AUC (area under the receiver operating characteristic curve) was used to identify the optimal logistic regression model (highest AUC) in each biomarker-defined subset. For multi-predictor analysis, 10-fold cross validation was used.

Results: 854 patients (301 pCRs, 35%) with FTV evaluations at T0, T1, and T2, HR/HER2 and RPS subtypes, and pCR outcomes were included. Numbers of patients and pCR rates in individual subtypes are listed in Table 1. Of FTV variables, percent change at T2 was selected for inclusion in almost all subtype specific optimal models except HR+/HER2+. FTV at T0 (pretreatment tumor volume) was included in triple negative, HER2-/Immune+, and HER2+/ BP-HER2_or_Basal models. Using the current HR/HER2-specific model, the highest AUC (0.74) was found in triple negatives and the lowest AUC (0.68) was in HR+/HER2+. Using the proposed RPS-specific model, the highest AUC (0.84) was found in HER2-/Immune-/DRD+ and the lowest AUC (0.59) was found in HER2+/BP-Luminal cohorts. Table 1 shows AUCs estimated using predictions generated by HR/HER2- versus RPS-specific models, in the full cohort and in individual HR/HER2 sub-cohorts. AUCs were improved when RPS-specific models were used in full and in HR+/HER2-, HR+/HER2+, and triple negative cohorts. No improvement was observed in the HR-/HER2+ cohort where 97% (72/74) were HER2+/BP-HER2_or_Basal.

Conclusion: Improved prediction of pCR was observed using the RPS-specific MRI model compared to the current HR/HER2-specific model. A new preRCB workflow is being developed to combine MRI-based prediction with core biopsy assessment to re-direct “good responders” to surgery earlier and more precisely based on a patient’s biological subtype.

View original