Development and Validation of a Multi-Omics Model Integrating MR Radiomics and Immune Scores for Prognostic Prediction in Locally Advanced Nasopharyngeal Carcinoma

Abstract

Despite the widespread use of the TNM staging system in nasopharyngeal carcinoma (NPC), current prognostic prediction remains suboptimal due to its inability to capture tumour heterogeneity and microenvironmental characteristics. This study aimed to develop a deep learning-based multi-omics model integrating radiomics features, immune scores and clinical characteristics to improve the prediction of 5-year progression in locally advanced NPC patients. This retrospective study included 262 locally advanced NPC patients from two centres (161 from Zhongnan Hospital and 101 from Tongji Hospital). MRI sequences (T1, T2, T1c) were pre-processed and registered. Tumour regions were automatically segmented using a pre-trained 3D-UNet model. Radiomics features were extracted and selected through univariate logistic regression, mRMR and LASSO methods. Clinical features were screened using univariate analysis, while immunological markers were analysed through multivariate logistic regression. The final combined model integrated clinical, immunological and radiomic signatures. All three constructed signatures demonstrated robust predictive capability (AUC > 0.7) across validation sets. The combined model achieved superior performance with AUCs of 0.961 in training, 0.844 in internal validation and 0.798 in external validation sets. Sensitivity and specificity reached 0.818 and 0.860, respectively, in internal validation. Decision curve analysis confirmed the highest clinical net benefit for the combined model across different threshold probabilities. This study developed a novel multi-omics model integrating radiomics, immune scores, and clinical features to predict LA-NPC prognosis. The model provides a non-invasive, cost-effective tool for clinicians to design personalised treatment plans, demonstrating significant clinical utility in both internal and external validation cohorts.