889P - CT-based radiomics models to predict progression in locally advanced head and neck cancer treated with definitive chemoradiation

Background

Definitive chemoradiation (ChRT) is the treatment of choice in both unresectable and larynx-preservation locally advanced head and neck cancer (LAHNSCC). Validated biomarkers are needed to optimize the prediction of cancer outcomes. Imaging biomarkers can improve patient risk stratification and help personalize therapy.

Methods

Clinical data and baseline CT scans from retrospective LAHNSCC patients treated with definitive ChRT were collected. The dataset was split into training (80%) and test (20%). A 5-fold cross-validation was performed in the training set. From the primary tumor, 108 radiomic features were extracted. Survival analysis models and classification models were developed to assess actuarial progression-free survival (PFS) and 5-year progression (Yes/No), respectively. For the former, performance was evaluated using IPCW, C-Index (CI) and AUC, and for the latter, AUC, sensitivity (S), specificity (Sp), and accuracy (Acc) were used. Feature importance was measured by the SHAP method. Independent analyses using radiomics features +/- clinical variables were performed. Patient risk stratification in two groups using survival analysis models was assessed through Kaplan–Meier curves.

Results

Final dataset included 171 LAHNSCC patients; 55%, 21% and 24% were stage IVA, IVB and III respectively. Disease progression at 5 years was observed in 53% of patients. To predict PFS, a random survival forest model including 4 radiomic features and TNM provided the best results, with AUC of 0.64 and CI of 0.66. The model allowed patient stratification into low and high risk of progression (log-rank p < 0.005, HR 1.23 95%CI:0.60-2.54). A XGBoost model with 12 radiomic features and 4 clinical variables (primary tumor site [oral cavity the most important], TNM, age and smoking) to predict 5-year progression provided the best performance. This model yielded an AUC of 0.74, a S of 0.53, a Sp of 0.81 and an Acc of 0.66.

Conclusions

A model trained with clinical and radiomic data to predict 5-year progression outperformed that trained only with clinical variables. This model yielded more robust results than the PFS model, whose metrics resulted insufficient to draw solid conclusions. Further validation is required.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

INCLIVA Biomedical Research Institute.

Disclosure

V. Agusti, A. Nogué Infante, A. Jiménez Pastor, F. Bellvis Bataller, A. Fuster-Matanzo, B. Soriano, J.P. Fernández, J.A. Molina, Á. Alberich Bayarri: Financial Interests, Personal and Institutional, Full or part-time Employment: QUIBIM SL. All other authors have declared no conflicts of interest.