S595
Clinical – Head & neck
ESTRO 2026
Purpose/Objective: This study aims to integrate multi-region radiomics features from radiotherapy target volumes and clinical parameters to construct an interpretable machine learning model for accurately predicting the 5-year recurrence risk in nasopharyngeal carcinoma (NPC) patients, and to explore the complementary value of primary tumor and lymph node features. Material/Methods: The study included 215 patients from public databases (randomly split 7:3 into training and validation sets) and 115 patients from our hospital (independent test set). Based on planning CT images, the primary tumor (GTVp), metastatic lymph nodes (GTVn), and fused targets (GTVpn) were delineated, and morphological, first-order statistical, and textural features were extracted. For GTVp features, LASSO regression (with 10-fold cross-validation to optimize parameter λ ) was first used to screen non-zero coefficient features from the full feature set. Six machine learning models (GBDT, MLP, LightGBM, AdaBoost, XGBoost, and GNB) were constructed, and model performance was evaluated using the area under the receiver operating characteristic curve (AUC). SHAP analysis was further employed to rank feature contributions and screen key features to reduce model complexity. Univariate and multivariate Cox regression analyses were performed on clinical variables to identify statistically significant indicators, which were integrated with radiomics scores to construct a prognostic nomogram. In the lymph node metastasis subgroup, single-target models (GTVp, GTVpn) and a dual-target combined model (GTVp + GTVn) were developed, and the optimal model was selected by comparing AUC values. Model validation was conducted using ROC curves, calibration curves, decision curve analysis, and net reclassification index (NRI) and integrated discrimination improvement (IDI) . Results: The GTVp-based GBDT model achieved AUCs of 0.889 in the training, validation, and test sets. After integrating clinical parameters, the nomogram model’s AUC improved to 0.898 (95% CI: 0.824-0.953), with significant improvements in NRI (0.315) and IDI (0.128). DCA showed clinical utility at threshold probabilities >0.2, and the calibration curve slope approached 1, indicating high consistency between predicted and actual risks. In the lymph node metastasis subgroup, the XGBoost model combining GTVp and GTVn features performed best (AUC= 0.918), significantly outperforming the single-target GTVp model (P < 0.05). The final multi-target fusion model achieved an AUC of 0.934 in the test set, with a calibration curve slope of 0.94 and an IDI of 0.183, demonstrating that dual- target features more comprehensively capture tumor heterogeneity. Conclusion:
Conclusion: High-risk RLN ENE is an independent adverse prognostic factor in NPC. N1/N2 patients with high-risk RLN ENE exhibit N3-like survival, supporting its potential incorporation into future refinements of N staging. References: [1] Pan J-J, Mai H-Q, Ng WT, Hu C-S, Li J-G, Chen X-Z, et al. Ninth Version of the AJCC and UICC Nasopharyngeal Cancer TNM Staging Classification. JAMA Oncol 2024. https://doi.org/10.1001/jamaoncol.2024.4354. Keywords: retropharyngeal lymph node; extranodal extension, Digital Poster 2358 Multiparameter Interpretable Machine Learning Model in Predicting 5-Year Recurrence of Nasopharyngeal Carcinoma: Based on Multi-Target Radiomics Weisi Wang, Chunsheng Wang, Mingjun Ding, Xia He Department of Radiation Oncology, Jiangsu Cancer Hospital & Jiangsu Institute of Cancer Research, Affiliated Cancer Hospital of Nanjing Medical University, Nanjing, China
Made with FlippingBook - Share PDF online