ESTRO 38 Abstract book

S28 ESTRO 38

978 patients remained in the analysis after visual inspection of the registration. The training and test datasets were well balanced for clinical parameters. The AUC for the elastic net model and the random forest model was 0.74 and 0.68, respectively. Both models identified the max dose to the right atrium (RA) and right coronary artery (RCA) as the most important factors associated with survival. The individual multivariate models (table 1) showed both RA and RCA were significant with a hazard ratio (HR) of 1.02, p=0.02. In both model’s (log) tumour volume was significant as well as N-stage. Interestingly, mean lung dose was not significant. The optimal cut point for the RA was 29.5Gy and for the RCA 15.5Gy. As these are neighbouring structures, the max dose was strongly correlated. The mean, 22.5Gy max dose, was applied to the test dataset and Kaplan Meier survival curves for the RA and RCA plotted, figure 1. The log rank for both curves was p<0.0001.

Conclusion In conclusion, a dose reduction from 70 Gy to 60 Gy (EQD 2 ) to the involved mediastinal lymph nodes in LA-NSCLC patients receiving (chemo)radiotherapy results in a lower incidence of severe toxicity and an increase OS and is therefore safe and promising. No increase in regional failures was observed. OC-0065 Cardiac dose and survival in lung cancer: which cardiac sub-structures matters most? A.McWilliam 1 , J. Khalifa 2 , E. Vasquez Osorio 1 , A. Abravan 1 , A. Marianne 1 , C. Faivre-Finn 1 , M. Van Herk 1 1 University of Manchester, Division of Clinical Cancer Sciences, Manchester, United Kingdom; 2 Institut Universitaire du Cancer de Toulouse, Oncopole, Toulouse, France Purpose or Objective There is increasing evidence that dose to the heart for non-small cell lung cancer (NSCLC) patients impacts overall survival. However, despite the growing body of evidence, cardiac sub-structures where excess dose has the greatest impact on overall survival have not been identified. In this work we implement two variable reduction techniques to identify these sub-structures and estimate radiotherapy dose-limits. Material and Methods 14 cardiac sub-structures were delineated on 5 template patients including: atria and ventricles, coronary arteries, valves and pulmonary arteries. 1,100 NSCLC patients were non-rigidly registered to the 5 template patients, mapping their planned radiotherapy dose to the template patient’s anatomy. Mean and max dose to each cardiac sub- structure were extracted. Across the 5 template patients, a mean of the mean dose was calculated and the max of the max dose extracted for analysis. The cohort was split into training (2/3 patients) and test (1/3 patients) datasets. Two variable reduction techniques were implemented: an elastic net LASSO model and a random forest survival model. Each model was optimised to extract the variables that contributed the most to overall survival. Models were tested on their ability to predict overall survival at 1 year in the test dataset, ROC analysis and area under the curve (AUC) calculated. The two most important variables, common to the two models, were selected for further analysis. Multivariate cox-proportional hazard models were created for each identified sub-structure. An optimal split on dose was performed for each variable with Kaplan Meier survival curves plotted for the test dataset to show significance. Results