Multimodal Data Integration to Predict Severe Acute Oral Mucositis of Nasopharyngeal Carcinoma Patients Following Radiation Therapy

Abstract

(1) Background: Acute oral mucositis is the most common side effect for nasopharyngeal carcinoma patients receiving radiotherapy. Improper or delayed intervention to severe AOM could degrade the quality of life or survival for NPC patients. An effective prediction method for severe AOM is needed for the individualized management of NPC patients in the era of personalized medicine. (2) Methods: A total of 242 biopsy-proven NPC patients were retrospectively recruited in this study. Radiomics features were extracted from contrast-enhanced CT (CECT), contrast-enhanced T1-weighted (cT1WI), and T2-weighted (T2WI) images in the primary tumor and tumor-related area. Dosiomics features were extracted from 2D or 3D dose-volume histograms (DVH). Multiple models were established with single and integrated data. The dataset was randomized into training and test sets at a ratio of 7:3 with 10-fold cross-validation. (3) Results: The best-performing model using Gaussian Naive Bayes (GNB) (mean validation AUC = 0.81 ± 0.10) was established with integrated radiomics and dosiomics data. The GNB radiomics and dosiomics models yielded mean validation AUC of 0.6 ± 0.20 and 0.69 ± 0.14, respectively. (4) Conclusions: Integrating radiomics and dosiomics data from the primary tumor area could generate the best-performing model for severe AOM prediction.

Keywords: acute mucositis; dosiomics; multimodal data integration; nasopharyngeal carcinoma; radiomics.

Figure A1

Schematic diagram of patient selection.

Figure A2

Dose maps of NPC patients undergoing IMRT. (a) 3D view of NPC patient. (b) DVH of multiple VOIs. (c). Axial view of patient in different VOIs. (d). Coronal view of patient. (e). Sagittal view of patient.

Figure 1

VOI examples for NPC patients with CECT examination. (a) Region of GTVnp (orange), axial view. (b) Region of GTVn (blue) and PTVn_70 Gy (red), axial view. (c) Region of PTVn_60 Gy (green), coronal view. (d) DVH curve of four VOIs.

Figure 2

Scheme of feature selection and modeling. Training and validation sets were separated before data analysis. The training set of data was used for feature selection. The validation set of data was used for model evaluation. To further manipulate the numerical and categorical data, reduce the interactions, and solve the collinearity problems, random forest (RF) selection was applicated for radiomics, dosiomics, and integrated data. Three linear or non-linear models were developed with independent validation data sets with selected features. The area under the curve (AUC) was set as the main evaluation method for the model performance.

Figure 3

10-fold validation AUC results for the test set. (a) The AUC plot of GNB model for the GTVnp_R_cT1 data set. (b) The AUC plot of GNB model for the GTVnp_RD data set. (c) The AUC plot of LR model for the C&R&D data set. (d) The heatmap of mean AUC results for all models.

Figure 4

Feature importance of SHAP for XGBoost model of GTVnp_RD. From the highest to the lowest level, the features are categorized in GLSZM, log sigma 60 mm 3D GLCM, original GLDM, GLDM, and log sigma 10 mm 3D GLCM.