Room: AAPM ePoster Library
Purpose: To use a deep neural network for survival analysis (DNNSurv) using intratumoral radiomics and dosiomics features in non-small cell lung cancer (NSCLC) patients treated with radiotherapy in RTOG 0617.
Methods: A total of 362 NSCLC patients had a planning CT image, a dose map and survival data. The cohort was randomly split into a training and a test dataset (n=255, 107) in a stratified fashion, maintaining the rate of prescription dose level (training/test: 52/50% and 48/50% for 60 and 74 Gy). A total of 319 features consisting of 47 clinical/dosimetric features plus 14 shape, 129 radiomics and 129 dosiomics texture features were extracted from clinical tumor volume. After a series of feature selection was performed using supervised principal component analysis and Boruta algorithm, a DNNSurv model was built that transforms observed survival times into a series of jackknife pseudo conditional survival probabilities used as a response variable in standard regression analysis. Prognostic accuracy of the DNNSurv model was evaluated on the test set by concordance index (c-index) for predicting overall survival (OS) and time-dependent ROC-AUC for 2-year survival (2YS) prediction, and compared with those of Cox proportional hazards (CoxPH), LASSO-regularized Coxnet (LASSO-Cox), and additive kernel support vector regression (AK-SVR) models. Kaplan-Meier analysis and log-rank test were performed using the test set with median stratification (derived from the training set) of predicted values for each model.
Results: For predicting OS, the DNNSurv model achieved the highest c-index of 0.662, followed by the AK-SVR, LASSO-Cox and CoxPH models attaining the c-indices of 0.643, 0.623 and 0.619. The same rankings of the models were found for 2YS prediction showing the AUCs of 0.650, 0.605, 0.573 and 0.570. Only the DNNSurv model yielded statistically significant difference between low- and high-risk groups (p=0.0003).
Conclusion: Pseudo-observations approach using DNNSurv may improve prediction performance for survival.
Funding Support, Disclosures, and Conflict of Interest: This project was supported by NCI grants, U24CA180803(IROC) and U10CA180868(NRG).