The purpose of this study is to investigate the ability of multivariate penalized regression models with tumor grade, age, receptor status and DCE- and DW-MRI parametric maps, obtained early in the course of therapy, to predict pathologic complete response (pCR) status at the time of surgery. DCE- and DW-MRI data were collected (3.0T MR scanner, Philips Healthcare, The Netherlands) before and after the first cycle of NAC from thirty-three patients with Stage II or III breast cancer who participated in an IRB-approved clinical trial. Details on data acquisition and longitudinal registration of DCE- and DW-MRI parametric maps can be found elsewhere¹. Pathological complete responses were confirmed at the time of surgery. There were 12 patients with pCR and 21 who had residual disease at surgery. Voxel-based information of the efflux rate constant K_ep from the DCE-MRI and apparent diffusion coefficient ADC from the DW-MRI was extracted via the redundancy analysis in which the most uncorrelated percentiles of K_ep and ADC values were chosen after fitting a series of multiple linear regression models for those voxels with increase in K_ep and decrease in ADC, respectively, from pre- to post-NAC². Logistic regression with ridge penalty was employed to accommodate a large number of predictors (e.g., voxel-level K_ep and ADC) compared to our sample size (n=33). The performance of the full model, i.e., voxel-level K_ep & ADC, the receptor status, age, and tumor grade (VRC model), was compared to the same model without the receptors’ status (VC model) in terms of overfitting-corrected³ AUC and Brier score that measures the accuracy of prediction. For comparison, the prediction power of the conventional measure RECIST (RC model) was also assessed. The statistical significance of the difference in overfitting-corrected AUCs and Brier scores was investigated by generating a bootstrap distribution of the difference with 300 replicates.In Table 1, the overfitting-corrected AUCs and Brier scores resulted from the three logistic ridge regression models are reported along with the corresponding 95% CIs. The overfitting-corrected AUC resulted from VRC model outperforms the other two models, i.e., VC and RC models. Moreover, VRC model achieves the smallest overfitting-corrected Brier score, indicating that VRC model outperforms the others. The differences in overfitting-corrected AUCs between VRC and VC, and between VRC and RC model are illustrated in Figure 1. The shaded area 0.057 and 0.05 in Figure 1 indicates how overfitting-corrected AUC resulted from VRC is smaller than that from VC and RC, respectively. The shaded area 0.097 and 0.053 in Figure 2 indicate how the overfitting-corrected Brier score resulted from VCR is larger than that from VC and RC, respectively. Although the 95% CI on each overfitting-corrected statistics is not statistically significant, Figure 1 and 2 strongly support that the overfitting-corrected AUC and Brier score of VRC are likely to outperform the corresponding statistics based on VC and RC model.The study supports incorporating voxel-level information along with hormone receptor information, to improve the ability to predict treatment response for breast cancer patients using early imaging time points.