2870 - Development of an Interpretable Machine Learning-Based Prediction Model for Long-Term Dysphagia Following Intensity-Modulated Radiotherapy in Nasopharyngeal Carcinoma
Presenter(s)
X. Wang1, L. Wang2, G. Liu3, Y. Xiang4, and Z. Zhao5; 1Sun Yat-sen University Cancer Center, Guangzhou, Guangdong, China, 2Sun Yat-sen University Cancer Center, State Key Laboratory of Oncology in South China, Collaborative Innovation Center of Cancer Medicine, Guangzhou, China, 3Sun Yat-Sen Memorial Hospital, Guangzhou, China, 4Department of Nasopharyngeal Carcinoma, State Key Laboratory of Oncology in South China, Collaborative Innovation Center for Cancer Medicine, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Sun Yat-sen University Cancer Center, Guangzhou, China, 5Sun Yat-Sen Memorial Hospital, Guangzhou, Guangdong, China
Purpose/Objective(s): Radiation-induced long-term dysphagia remains a critical concern in locoregionally advanced nasopharyngeal carcinoma (LA-NPC) patients treated with intensity-modulated radiotherapy (IMRT). This study aimed to develop an interpretable XGBoost machine learning (ML) model integrated with SHapley Additive exPlanations (SHAP) to identify dosimetric predictors of long-term dysphagia and provide recommendations for dose limitations in clinical practice.
Materials/Methods: This single-center retrospective study included 421 LA-NPC patients treated with induction chemotherapy and IMRT between 2017–2019. Swallowing organs at risk (SWOARs: oral cavity, superior constrictor muscle [SCM], middle constrictor muscle [MCM], inferior constrictor muscle [ICM], supraglottic larynx [SGL], and glottic larynx [GL]) were defined and delineated. Patients’ clinical factors and dose-volume parameters of the SWOARs were retrieved from medical records and the IMRT treatment planning system, respectively. Long-term dysphagia (=5 years post-treatment) was assessed using the RTOG/EORTC scale and MD Anderson Dysphagia Inventory (MDADI). A least absolute shrinkage and selection operator (LASSO) logistic regression model, univariate and multivariate analysis were used to screen potential risk factors. XGBoost was employed for machine learning model construction and SHAP values were utilized to interpret the model. ROC curve analysis, calibration curve analysis, clinical decision curve analysis, sensitivity, specificity, accuracy, and F1 score were used for evaluating the model’s performance.Dose-effect analysis was performed to determine dose limitations.
Results: With a median follow-up of 73.7 months, 24.5% of patients developed long-term dysphagia, with 3.8% experiencing severe long-term dysphagia (RTOG grade 2-3). Mean dose of inferior constrictor muscle (ICMDmean), volume 55 Gy of oralcavity (OralCavityV55), volume 60 Gy of the middle constrictor muscle (MCMV60) and dose 90% of the middle constrictor muscle (MCMD90) were the most important predictors for long-term dysphagia. The prediction model based on these factors demonstrated excellent discriminative abilities (AUC: train: 0.995, 95 %CI: 0.989-1; test: 0.978, 95 % CI: 0.957–0.998). Mean dose of middle constrictor muscle (MCMDmean) was identified as the independent predictor for severe long-term dysphagia. ICMDmean<44Gy, OralCavityV55<11% , MCMV60<27% and MCMD90<48Gy for long-term dysphagia and MCMDmean<63Gy for severe long-term dysphagia are suggested as rational dose limitations.
Conclusion: The XGBoost machine learning model constructed by comprehensive analysis can be used to predict the risk of long-term dysphagia after IMRT in LA-NPC patients. Additionally, rational dose limitations were suggested. The results of the ML model can be better interpreted in depth by combining SHAP method.