C. Liu, J. Gong, and M. Shi; Department of Radiation Oncology, First Affiliated Hospital of Air Force Medical University, Xian, China
Purpose/Objective(s):Distant metastasis (DM) is the main failure pattern of nasopharyngeal carcinoma (NPC) treatment. Prediction model that accurately identify patients with genuine high DM risk is demanded. Here we identified clinical risk factors and radiomic features and constructed a prediction-score model for DM prediction in non-endemic nasopharyngeal carcinoma. Materials/
Methods: 510 patients with non-metastatic NPC from our institution and 60 patients from other institutions were enrolled and followed for at least 3 years. 25 Clinical features were analyzed for screening clinical risk factors. Pre-treatment contrast enhanced CT images were collected and 1316 radiomic features were extracted from tumor by an open source software. 408 patients from our institution were randomly assigned to training cohort and the rest 102 were assigned to internal validation cohort. 60 outside patients were assigned to external validation cohort. Clinical risk factors were determined by univariate and multivariable analysis. The least absolute shrinkage and selection operator (LASSO) Cox regression model was applied to select the most predictive radiomic features. Prediction models with clinical risk factors and radiomic signature respectively and the combination were developed and validated. The Harrell Concordance Index (C-index) and the time-dependent AUC were applied to evaluate the model’s efficiency. Results: There were 153 of 570 patients who had distant metastasis. Six clinical features, including age, KPS score, N stage, AJCC stage, hemoglobin and predictive nutritional index (PNI) before treatment, were screened out to develop the clinical model. Twenty-two radiomics features were selected to develop the radiomic signature. The final nomogram, which included the six clinical features and radiomic signature, achieved satisfactory discriminative performance and outperformed the clinical or radiomic signature alone models for predicting distant metastasis. C-index for combined, radiomic and clinical models were 0.759 vs. 0.723 vs. 0.679in training cohort and 0.727 vs. 0.711 vs. 0.610 in internal validation cohort and 0.692 vs. 0.647 vs. 0.594 in external validation cohort. Patients were stratified by the nomogram into low and high-risk groups with different DM risk (cut-off value: 0.035). Patients with low risk had better DMFS than with high risk in training (p < 0.001; 3-y DMFS 93.4% vs. 72.6%), internal (p < 0.001; 3-y DMFS 92.5% vs 63.3%) and external validation cohorts (p < 0.05; 3-y DMFS 88.7% vs 67.5%). 3-year AUC of combined, 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 cinical-radiomics prediction model consisting of six clinical features and radiomic signature that can evaluate DM risk of non-endemic NPC patients with high efficiency. Such model might aid in risk-adapted treatment decisions and personalized follow-up strategies.
