2794 - Development and Validation of a Hybrid Clinical-Delta Radiomics Model to Predict Distant Metastasis after Definitive Radio-Chemotherapy in Non-Endemic Nasopharyngeal Carcinoma

Purpose/Objective(s): Distant metastasis (DM) is the main failure pattern of nasopharyngeal carcinoma (NPC) treated by radical chemoradiation despite of satisfying local control. Currently widely used AJCC staging could not precisely predict DM risk. It is of great value to discover more accurate prediction parameters that distinguish patients with genuine high DM risks. Besides, currently published DM prediction research mainly focused on patients from endemic area, making these results inapplicable to non-endemic patients. Here we identified clinical risk factors and delta radiomic features and constructed a prediction-score hybrid model for DM prediction in non-endemic nasopharyngeal carcinoma.

Materials/Methods: 510 non-metastatic NPC patients from our institution and 100 patients from two outside institutions were enrolled and followed for at least 3 years. 25 clinical features were screened for risk factors. Contrast enhanced CT images were collected and 1316 radiomic features were extracted. Metastatic lymph nodes with diameter larger than 3 cm were analyzed for delta radiomics features. 408 patients were randomly assigned to training cohort and the rest 102 to the internal validation cohort. 100 patients from outside institutions were assigned to external validation cohort. Clinical risk factors were determined by univariate and multivariable analysis. The LASSO Cox regression model was applied to select the most predictive features. Prediction models with clinical risk factors and delta radiomic signatures respectively and the combination were developed and validated. The C-index and the time-dependent AUC were applied to evaluate the model’s efficiency.

Results: There were 153 patients who had distant metastasis. Five clinical features, including age, KPS score, N stage, AJCC stage, and hemoglobin were screened out to develop the clinical model. Twenty-nine radiomics features were selected to develop the delta radiomic signature. The final nomogram, which included above mentioned signatures, achieved satisfactory discriminative performance and outperformed the clinical or radiomic signature alone models for predicting distant metastasis. C-index for hybrid, radiomic and clinical models were 0.774 vs. 0.753 vs. 0.682 in training cohort and 0.747 vs. 0.731 vs. 0.640 in internal validation cohort and 0.706 vs. 0.667 vs. 0.594 in external validation cohort. Patients were stratified by the nomogram into low and high-risk groups with different DM risk. Patients with low risk had better DMFS than with high risk in training, internal and external validation cohorts. 3-year AUC of hybrid, radiomic and clinical models were 0.750 vs 0.768 vs 0.704. The calibration curves showed excellent agreement between the predicted and actual DMFS.

Conclusion: We developed a Hybrid Clinical-Delta Radiomics Prediction Model that predict DM risk of non-endemic NPC patients with high efficiency. Such model might aid in risk-adapted treatment decisions and personalized follow-up strategies.