2159 - Application of Radiomic and Dosiomic Analyses to Predict for Pneumonitis in Patients with Locally Advanced Non-Small Cell Lung Cancer on the NRG Oncology RTOG 0617 Dataset

T. Upadhaya, E. McKenzie, S. C. Zhang, I. J. Chetty, and K. M. Atkins; Department of Radiation Oncology, Cedars-Sinai Medical Center, Los Angeles, CA

Purpose/Objective(s): We performed radiomic and dosiomic analyses to predict radiation pneumonitis (RP) in patients with locally advanced non-small cell lung cancer (NSCLC) treated with chemo-RT. Materials/

Methods: NRG Oncology/RTOG 0617 data from the NCTN were analyzed. Radiomic and dosiomic features were extracted from pre-treatment planning CT’s and 3D dose distributions respectively, using Image Biomarker Standardization Initiative compliant software. Analysis was centered on normal lung tissue contours (both lungs-GTV) to predict RP>grade 2 (CTCAE). ROC and correlation analyses were used to identify independent predictors. Models were evaluated based solely on clinical, dosiomic, and radiomic features along with combined feature, ensemble models. Models including support vector machine, random forest, and LASSO/GLM were trained using 5-repeat 10-fold cross-validation. Independent predictors were ranked on importance score and stepwise-forward feature selection was used to identify the subset of features minimizing validation error. Synthetic Minority Oversampling Technique (SMOTE) was used to minimize effects of imbalanced class sizes. The model with highest AUC was applied to unseen test patients.
Results: 451 patients (315 for training and 136 hold-out for testing) were available for analyses and included clinical features, dose, and toxicity information. RP>Grade2 occurred in 15% (n=67) with similar rates observed between high (74 Gy) and low (60Gy) dose arms. Among statistically significant features (AUC>0.5), 68 radiomics, 54 dosiomics, and all 4 clinical features (AJCC stage, chemo, age, Zubrod score) were independent predictors of RP (correlation coefficient <0.8). Models based on clinical, dosiomic, and radiomic features independently achieved AUC’s of 0.54, 0.62, 0.67 on validation and 0.52, 0.62, 0.63 on testing datasets, respectively. An ensemble model based on combination of clinical, DVH, dosiomic or radiomic features yielded AUC’s ranging from 0.58-0.68 (validation) and 0.57-0.62 (testing). Among all ensemble models evaluated, the combined radiomic and dosiomic model outperformed other models with AUC of 0.68 (testing).
Conclusion: Results are suggestive that radiomic features on planning CT’s provide complimentary information to enhance model performance implying that such underlying features on pre-treatment CT may serve as a biomarker for prediction of post-treatment RP. Ensemble modeling combing clinical, dosiomics, and radiomics features yielded the best performance. Conclusive results cannot be drawn due to the limited sample size despite the use of statistical approaches to minimize bias. Future investigation using lung substructure-based analyses is warranted to provide more focused associations between regions of lung/dose/volume and pneumonitis outcomes.