OBJECTIVE: This study aims to deeply mine the radiomic features of the primary tumor (GTVp) and cervical metastatic lymph nodes (GTVn) in nasopharyngeal carcinoma (NPC) to construct a multi-target synergistic prediction model. The goal is to achieve precise risk stratification for 5-year recurrence in patients at the initial diagnosis stage, providing a scientific basis for formulating individualized treatment and follow-up strategies. METHODS: A total of 330 NPC patients were retrospectively included and divided into a training set (n = 150), an internal validation set (n = 65), and an independent external validation set (n = 115). Three target construction strategies-primary tumor alone (GTVp), fused target volume (GTVpn), and multi-target combination (GTVp&n, independent feature extraction followed by merging)-were employed. Core features were selected through dimensionality reduction using Least Absolute Shrinkage and Selection Operator (LASSO) regression combined with 10-fold cross-validation. Six machine learning algorithms, including XGBoost, GBDT, and MLP, were compared to determine the optimal scheme. Using SHAP interpretability analysis and multivariate logistic regression, a combined prognostic nomogram was constructed by integrating clinical factors and the radiomics score (R-score), and its generalizability was evaluated on the external validation set. RESULTS: For the single GTVp target, 21 radiomic features were selected via LASSO, among which the GBDT model performed best (training set AUC = 0.889). After streamlining to 6 features via SHAP analysis, the combined clinical nomogram yielded AUCs of 0.898, 0.879, and 0.857 in the training, internal validation, and external validation sets, respectively. Subgroup analysis of patients with lymph node metastasis showed that the XGBoost model using the GTVp&n strategy (independent feature extraction followed by merging) achieved the highest efficacy (training set AUC = 0.918), which was significantly superior to the single GTVp model (P = 0.035). The performance of the upgraded combined model was further enhanced, with the training set AUC reaching 0.934 and the external validation set AUC increasing to 0.879. Quantitative indicators (IDI = 0.139, NRI = 0.201) confirmed that incorporating lymph node heterogeneity features plays a key role in improving model accuracy and generalizability. CONCLUSION: This study developed and validated a synergistic radiomics model integrating information from both the primary tumor and metastatic lymph nodes. By simultaneously capturing the "source information" of the primary tumor and the "evolutionary information" of the lymph nodes, this synergistic model showed improved predictive performance compared to single-target models.. It offers potential quantitative evidence that may assist in formulating individualized intensive treatment strategies for NPC.
Wang et al. (Tue,) studied this question.