A Multimodal Machine Learning Model Combining Dosimetric and Radiomic Shape Features for Predicting Moderate‑to‑Severe Xerostomia After Radiation Therapy for Nasopharyngeal Carcinoma.

PURPOSE: Radiation-induced xerostomia remains a frequent and debilitating complication in patients with nasopharyngeal carcinoma undergoing radiation therapy, severely impacting quality of life. Accurate prediction of xerostomia severity is crucial for personalized treatment planning. This study aimed to develop and validate a machine learning model that integrates dosimetric parameters with radiomic shape features from planning computed tomography images to predict moderate‑to‑severe xerostomia.

METHODS AND MATERIALS: In this retrospective study, we included 100 patients with nasopharyngeal carcinoma treated with intensity modulated radiation therapy. Radiomic shape features were extracted from the overlap volume between the planning target volume and the left parotid gland. These were combined with standard dosimetric parameters (eg, V20 Gy and V30 Gy). A rigorous correlation-based feature selection was implemented to address multicollinearity. A random forest classifier was developed to predict xerostomia grade (grade 1 vs ≥grade 2), and its stability was rigorously assessed through 100 iterations of random train-test splitting. Model performance was evaluated using the area under the receiver operating characteristic curve, accuracy, and confusion matrix analysis.

RESULTS: The model demonstrated moderate performance with modest variability in predicting moderate‑to‑severe xerostomia, achieving a mean area under the receiver operating characteristic curve of 0.78 (SD, 0.09) and an accuracy of 0.71 (SD, 0.09) across 100 random train‑test splits. Feature selection stability analysis identified 6 consistently important predictors: "Elongation," "Maximum2DDiameterSlice," "V20," "Flatness," "LeastAxisLength," and "MajorAxisLength." Radiomic shape features predominated among the most frequently selected and most influential parameters. Robust importance quantification further identified "V30," "MeshVolume," and "Sphericity" as the top contributors to the model's predictions.

CONCLUSIONS: This study suggests a robust machine learning framework for predicting moderate‑to‑severe xerostomia, indicating the complementary value of radiomic shape features to traditional dosimetric metrics. The integration of these features may contribute to more comprehensive toxicity risk assessment and warrants further investigation in clinical decision-making contexts.