Development and validation of an MRI radiomics-based interpretable machine learning model for predicting the progression-free survival in locally advanced nasopharyngeal carcinoma.

Background: Locally advanced nasopharyngeal carcinoma (LANPC) is a common malignant tumor of the nasopharynx, characterized by poor prognosis and a high susceptibility to recurrence and metastasis after surgery. The aim of this study was to establish and validate a radiomics model based on clinicopathological data and magnetic resonance imaging (MRI) information to predict progression-free survival (PFS) in LANPC patients, and to reveal the internal prediction process of the model through SHapley Additive exPlanation (SHAP) and image visualization techniques. Methods: A total of 1,098 patients with pathologically and clinically diagnosed LANPC were recruited from three hospitals {training, n=700 [70% from hospitals I (Guangxi Medical University Cancer Hospital) and II (Wuzhou Red Cross Hospital)]; internal validation, n=300 (remaining 30%); and external validation, n=98 [hospital III (The Second Affiliated Hospital of Guangxi Medical University)]}. Univariate Cox analysis and least absolute shrinkage and selection operator (LASSO) Cox regression were used to select radiomics features. A combined model integrating the radiomics score (radscore) and important clinicopathological factors was constructed using the eXtreme Gradient Boosting (XGBoost) algorithm. SHAP and image visualization techniques were used for interpretability analysis of prognostic models. Results: Optimal predictive performance was observed in the combined model, which integrated the radscore and important clinicopathological factors [induction chemotherapy (IC), Epstein-Barr virus (EBV)-DNA, and albumin], with Harrell concordance index (C-index) values of 0.762, 0.729, and 0.752 in the training, internal, and external validation cohorts, respectively. Ten radiomics features with the highest predictive contributions were identified using the SHAP algorithm. Two LANPC patients with similar clinicopathological stages but distinct risk levels were selected to visualize the top three radiomics features, revealing notable pixel-level visual differences in the largest layer of the tumor images. Kaplan-Meier survival analysis revealed prognostic differences between low- and high-risk groups, and the model's performance was stable across subgroups (all log-rank P<0.05). Conclusions: The interpretable model was able to accurately predict the PFS in LANPC patients. SHAP and image visualization techniques provided quantitative contribution values and image-level radiomics information, which could provide valuable additional information for individualized prognostic evaluation and clinical decision-making for patients with LANPC.