Abstract
PURPOSE: This study developed and validated a multimodal machine learning model integrating clinical, radiomics, and dosiomics characteristics to predict radiation-induced temporal lobe injury (RTLI) in nasopharyngeal carcinoma patients. METHODS: This study utilized dual-center retrospective cohorts comprising 213 patients. From pre-treatment 3D computed tomography (CT) scans and radiation dose maps, we extracted 43 radiomics features and 43 dosiomics features. These quantitative imaging features underwent a rigorous three-stage selection pipeline-including auto-correlation analysis, significance testing (t-tests), and least absolute shrinkage and selection operator-regularized logistic regression (LASSO-LR)-yielding 24 optimal image-derived predictors (13 radiomics and 11 dosiomics features). In parallel, eight clinical features were preserved in all final models due to their established clinical relevance and prior biological significance. We systematically evaluated twelve predictive configurations by integrating three machine learning algorithms-Support Vector Machine (SVM), Random Forest (RF), and Logistic Regression (LR)-with four feature modalities: clinical-only, radiomics-only, dosiomics-only, and an integrated feature set. RESULTS: The SVM-based integrated model demonstrated superior predictive performance, achieving an AUC of 0.819 during internal validation while maintaining robust generalizability in external testing (AUC = 0.714). The model exhibited excellent specificity (0.844) and remarkable cross-cohort stability, as evidenced by low performance variability (standard deviation: ±0.077 in validation, ±0.056 in testing). These findings highlight the synergistic value of integrating clinically established parameters with quantitative characterization of anatomical heterogeneity (radiomics) and radiation dose distribution patterns (dosiomics) for accurate prediction of radiation-induced temporal lobe injury. CONCLUSION: This clinically applicable predictive framework enables early identification of high-risk RTLI patients, facilitating timely interventions to mitigate neurotoxicity. The model's demonstrated multicenter validity, high specificity, and balanced performance across evaluation metrics address critical limitations of current prediction approaches, representing a significant advance in personalized radiation oncology for nasopharyngeal carcinoma patients.