Development of CT-based radiomic model to predict 5-year progression-free survival (PFS) in locally advanced head and neck squamous cell carcinoma (LAHNSCC) treated with definitive chemoradiation.

6076 Background: Definitive chemoradiation is the current standard treatment in LAHNSCC. Prognostic biomarkers that enable to reliably predict LAHNSCC disease progression are needed to stratify the risk of progression to tailor treatment intensity in future clinical trials. Imaging biomarkers have emerged as promising alternatives to non-invasively predict patient outcomes. Aim: to identify an imaging signature to predict progression-free survival (PFS) in LAHNSCC. Methods: A single-center, retrospective, observational study was designed. Clinical data and pre-treatment CT images from LAHNSCC patients treated with definitive chemoradiation from 2014 to 2022 were collected. Manual tumor segmentation was performed slice by slice by a qualified image technician and supervised by a radiologist with 20+ years in HNSCC using the DICOM viewer of Quibim platform. A total of 108 radiomic features (shape and textures) were extracted from each region of interest. Feature reduction techniques were applied to select the characteristics used in the classification model. The primary endpoint of this study was 5-year PFS. Random Forest, Support Vector Classifier (SVC), Logistic Regression, Gradient Boosting (GB) and Extreme Gradient Boosting classification models were used. A 5-fold cross validation strategy was followed to evaluate the performance of the models in terms of AUROC, sensitivity, specificity, and accuracy. The importance of each feature in the model was measured by applying the Shapley Additive exPlanations (SHAP) method. Results: Baseline CT exams from 102 LAHNSCC patients were included; 50% and 23% were stage IVA and IVB, respectively. The remaining 27% (stage II [3%] and stage III [24%]) were treated in an organ-preservation intention. 67% of patients presented locoregional or distant progression at 5 years. All radiomic features (108) and 6 clinical variables were used in the predictive model, for which a cross-validation was performed. GB as the feature selection technique and SVC as the classification model provided the best results. The final model was able to predict 5-year PFS with an average AUC of 0.82 (95% CI: [0.73, 0.91]), a sensitivity of 0.69 (95% CI: [0.57, 0.81]), a specificity of 0.82 (95% CI: [0.7, 0.94]) and an accuracy of 0.73 (95% CI: [0.62, 0.84]). From most to least important, features contributing to the predictive model were: TNM stage (SHAP: 0.038), glszm Small Area Emphasis (SHAP: 0.026), glcm Idmn (SHAP: 0.023), glszm Large Area High Gray Level Emphasis (SHAP: 0.018) and glrlm Run Entropy (SHAP: 0.015). Conclusions: Radiomic biomarkers from pre-treatment CT images in combination with TNM stage were predictive of 5-year PFS to chemoradiation in LAHNSCC patients and might be helpful for patient risk stratification. Further validation of these imaging biomarkers is ongoing.