Predictive Modeling of Acute Radiation-Induced Dermatitis in Nasopharyngeal Carcinoma Patients Undergoing TomoTherapy Using Machine Learning with Multimodal Data Integration

Jiabiao Hong¹Yuhao Lin¹Xiaoting Lin¹Linghui Yan¹Jihong Chen¹Huabing Chen²Miaomiao Zeng^2*Shuzhen Yuan^3*

• ¹Department of Radiation Oncology, Clinical Oncology School of Fujian Medical University, Fujian Cancer Hospital, Fuzhou 350014, Fujian, China, Fuzhou, China
• ²The First Affilated Hospital, Hengyang Medical School, University of South China, Hengyang 421001, Hunan, China, Hunan, China
• ³Department of Sport Medicine, Southern Hospital Ganzhou Hospital (Ganzhou People's Hospital), Ganzhou 341000, Jiangxi, China, Jiangxi, China

Purpose: Radiation dermatitis(RD) is a common and debilitating side effect of radiotherapy in nasopharyngeal carcinoma(NPC) patients. Traditional predictive models lack sufficient accuracy for assessing acute radiation dermatitis(ARD) after tomotherapy treatment. This study aims to integrate clinical, dosimetric, and radiomic features to enhance the accuracy and robustness of predictions, thereby promoting a more personalized risk assessment for NPC patients undergoing tomotherapy. Methods: A cohort of 161 NPC patients who underwent Tomotherapy was retrospectively analyzed. Clinical, dosimetric, and radiomic features were extracted for the purpose of model development. Feature selection was conducted using statistical tests and Least Absolute Shrinkage and Selection Operator(LASSO) regression. Several machine learning algorithms were then employed to construct the predictive models, including Logistic Regression, Support Vector Machine(SVM), K-Nearest Neighbors(KNN), Random Forest, Extra Trees, XGBoost, Light Gradient Boosting Machine (LightGBM), and Multilayer Perceptron(MLP). These models were built based on clinical, radiomic, dosiomic, and combined feature sets. Model performance was assessed by evaluating the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. To ensure fairness in comparisons, five-fold cross-validation was applied during the training of all models in the training cohort. Results: The combined model, integrating clinical, radiomic, and dosiomic features, demonstrated the highest predictive accuracy, achieving an AUC of 0.916 (95% CI: 0.866–0.967) in the training cohort and 0.797 (95% CI: 0.616–0.978) in the validation cohort. In comparison, the clinical model (AUC = 0.704), radiomic model (AUC = 0.865), and dosiomic model (AUC = 0.640) had lower predictive performance. SVM method demonstrated superior overall performance across various model constructions. The combined model based on the SVM method exhibited optimal predictive performance, achieving the best results in both the test and validation cohorts. Conclusions: The developed combined prediction system achieves superior performance in anticipating severe ARD in NPC undergoing tomotherapy cases. This tool facilitates pre-therapeutic risk stratification, dosimetric parameter refinement, and evidence-based scheduling of preventive skin management protocols, offering a paradigm-shifting approach to individualized cutaneous protection strategies.