A Multimodal fusion model for predicting nasopharyngeal necrosis after re-irradiation in recurrent nasopharyngeal carcinoma.

OBJECTIVES: To construct a multimodal model integrating computed tomography (CT) imaging, radiotherapy dose, and clinical features to address the clinical challenge of accurately predicting the risk of nasopharyngeal necrosis following re-irradiation for recurrent nasopharyngeal carcinoma (NPC). METHODS: This retrospective study included 126 patients with recurrent NPC (44 with necrosis and 82 without necrosis). Clinical baseline characteristics, radiomic features from planning CT images, and three-dimensional radiotherapy dose-derived radiomic features were collected and extracted. A two-stage feature selection strategy was employed to identify relevant features. Three single-modality and four multimodal fusion machine learning models were developed based on the selected features. Model performance was evaluated using internal (n = 26) and external (n = 9) test sets, with assessment metrics including the area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value, and negative predictive value. RESULTS: The multimodal fusion model demonstrated overall superior performance compared to all single-modality models. The CT-dose fusion model achieved AUCs of 0.869 and 0.972 in the internal and external test sets, respectively, outperforming the single CT, single-dose, and single clinical models. CONCLUSION: An artificial intelligence model integrating CT radiomic features and radiotherapy dose-derived characteristics demonstrated robust performance in predicting nasopharyngeal necrosis following re-irradiation for recurrent NPC, exhibiting stable predictive ability and strong generalizability. This model provides a powerful tool for the early identification of high-risk patients, with considerable promise for clinical application.