Relationship between Cardiac Dose and Overall Survival after Radiotherapy in a Large Cohort of Patients with Locally Advanced NSCLC

Purpose/Objective(s)

While correlations between cardiac dose and outcomes in NSCLC patients have been reported in several retrospective studies, results remain, nonetheless, inconclusive, with different dosimetric parameters being reported in different studies. This could be due to increased radiosensitivity in only some cardiovascular substructures, inconsistent structure contouring, and the lack of strong validation methodologies. In this study, we analyzed the relationship between cardiac dose and overall survival (OS), while trying to tackle these shortcomings and using a multivariate analysis on a large, multi-institutional dataset.

Materials/Methods

Our dataset consisted of 953 stage-III NSCLC patients treated with 66 Gy in 24-33 fractions with concurrent, sequential, or no chemotherapy and using IMRT or VMAT. Data was obtained retrospectively from patient files. Consistent and accurate contours of the heart, cardiac chambers, and great vessels were obtained automatically using our recently developed deep learning model. For the survival analysis, we trained and validated a multivariate Cox regression using a 5-fold cross validation with an inner development set (20%). For the model, a fixed number of non-correlated parameters were chosen at random. This random selection was repeated 1000 times and the set of parameters with the highest c-index in the training set was kept. PTV was always included. This process was repeated for each split with successively increasing number of parameters. The c-index peaked with limited overfitting in the development set using 4 parameters. We report the results in the test set using the 4 parameters chosen in each split.

Results

Median OS was 25 months (95 % CI: 22-27) and median follow-up was 42 months (95 % CI: 40-46). The multivariate Cox regression models with 4 parameters achieved an average c-index of 60.95 ± 1.78 % over the test sets (Table 1). All parameters chosen for the 5 models had statistically significant hazard ratios (95% CI) and, in all five, cardiovascular dosimetric parameters were chosen. In two of the splits, dosimetric parameters associated with the superior vena cava (SVC) were selected, V60 (p<0.001) and D80 (p=0.037) and in three splits whole heart parameters were chosen: Heart D35 (p=0.011), Heart D25 (p=0.004), and Heart V5 (p<0.001).

Conclusion

We found strong evidence that dose to cardiovascular substructures affects OS of NSCLC patients in multivariate analysis. Specifically, we found SVC and whole-heart dosimetric parameters consistently associated with OS. We further recommend reducing whole heart and superior vena cava dose in treatment of NSCLC.