Predicting treatment effect on distant recurrence free survival from functional tumor volume change during neoadjuvant therapy: A Bayesian hierarchical model of I-SPY 2 MRI and survival data.

618Background: In neoadjuvant cancer trials, early endpoints that predict treatment effect on survival identify promising agents early and support accelerated regulatory approval. However, binary endpoints like pathologic complete response inadequately characterize the full distribution of residual disease in breast cancer, while promising continuous biomarkers like MRI derived functional tumor volume (FTV) are associated with survival outcomes but not widely collected. Using a Bayesian hierarchical model, long-term treatment effects on distant recurrence free survival (DRFS) can be predicted from continuous MRI-derived functional tumor volume (FTV) in the I-SPY 2 platform trial. Methods: I-SPY 2 treated 2117 patients from 2010-2022 (12 weeks of paclitaxel ± experimental agent followed by 4 cycles of doxorubicin + cyclophosphamide), and 1, 859 underwent dynamic contrast-enhanced MRI at baseline and after 12 weeks of neoadjuvant therapy. MRI-derived functional tumor volume provided volumetric quantification of dynamic tissue enhancement. ΔFTV was defined as the ratio of 12-week to baseline FTV. A Bayesian joint hierarchical model (brms) fit treatment effects on ΔFTV and DRFS for each treatment regimen by HR/HER2 subtype, controlling for clinical nodal status, clinical T stage, grade, and calendar year. Arms with < 8 subjects are excluded. Performance was assessed using cross-validation, predicting DRFS treatment effects in one held out fold at a time from the learned ΔFTV-DRFS association in the remaining data, then comparing predicted to actual DRFS treatment effect. Sensitivity analyses on priors will be presented. Results: Across 1753 patients and 45 treatment–subtype combinations, the estimated treatment effects on ΔFTV and DRFS were highly correlated (posterior correlation -0.91; 95% CrI -1.00 to -0.23). Predicted DRFS treatment effect from ΔFTV demonstrated strong concordance with actual DRFS treatment effects (Pearson r = 0.94 in TNBC; 0.97 HER2+; 0.80 HR+HER2-). The top 5 treatment-subtype regimens ranked by predicted and actual DRFS were identical. 20 regimens predicted to have > 70% probability of DRFS benefit over subtype specific controls showed DRFS improvement, yielding 100% specificity and 69% sensitivity at this decision threshold. Conclusions: We demonstrate internally validated prediction of neoadjuvant treatment effect on DRFS from MRI-derived change in functional tumor volume in the I-SPY 2 trial of high-risk early breast cancer. This suggests continuous imaging measures capture a range of response to therapy while Bayesian approaches can be effective for predicting treatment effects. This encourages collecting MRI biomarkers in trials to facilitate validation as early endpoints supporting decisions in screening platform trials as well as regulatory accelerated approval.