Validation and Optimization of a Cardiac NTCP Model for Breast Cancer Patients Based on Cardiac Substructures and 3-dimensional Dose Distributions

Abstract

Purpose/Objective(s): A dose-effect relationship between mean heart dose (MHD) and acute coronary event (ACE) rate was reported in a study by Darby et al. (N Engl J Med 2013; 368:987-98). The increase was 7.4% per Gy MHD for the entire follow up period and w16% per Gy in the first 9 years. However, the main limitations of that study were the study design (case control study) and the use of reconstructed MHD. The first aim of this study was to validate the Darby et al. results with planned dose data from a cohort of breast cancer (BC) patients. Secondly, we investigated the potential for optimization of the NTCP model by replacing the MHD with dose-volume parameters of cardiac substructures (left and right atria and ventricles). Materials/Methods: The cohort consisted of 910 consecutive female BC patients treated with postoperative radiotherapy after breast conserving surgery. The primary endpoint was the occurrence of an ACE (defined according to Darby et al.) with up to 9 years follow-up. For each patient, both the MHD and various dose-volume histogram (DVH) parameters of the cardiac substructures were collected from the 3D planning data calculated using CT scans. External validation of the Darby et al. model was performed by testing its performance in terms of calibration (Hosmer and Lemeshow (HL) test) and discrimination (C-statistic). Cox multivariate regression analysis was used to investigate improvement of the model by using DVH parameters other than the MHD. A bootstrap method was used for internal validation. Results: The median follow-up time was 7.6 years (range: 0.1 -10.1). The median MHD was 2.35 Gy (range: 0.51 to 15.25). The 9-year cumulative incidence of ACE was 3.9%. The cumulative incidence of ACE increased by 15.8% per Gy (P = 0.034) within the first 9 years of follow-up. This NTCP-model performed well with a C-statistic of 0.71 (95% CI = 0.61 -0.80), with good agreement between observed and expected events (HLtest: P = 0.406). Additional analysis showed that the V5 of the left ventricle (LV-V5) was most important. By replacing the MHD with the LVV5, model performance improved (C-statistic: 0.74; 95% CI = 0.65 -0.82). Replacement of the dichotomous variable with the variable weighted sum of the risk factor(s) per patient and age, further improved model performance (C-statistic: 0.84; 95% CI = 0.75 -0.92). Conclusion: We externally validated the Darby et al. model in a prospective cohort study using clinical 3D dose distribution data. The relative increase of 15.8% per Gy MHD is similar to that observed by Darby et al. within the first 9 years. Model performance can be significantly improved by replacing the MHD by the LV-V5 and by using the weighted sum of the risk factor(s) per patient and age. This study confirms the importance of reducing the dose to the heart.