A Machine Learning–Based Prediction Model for Dysphagia After Radiotherapy in Nasopharyngeal Carcinoma Survivors

A Machine Learning–Based Prediction Model for Dysphagia After Radiotherapy in Nasopharyngeal Carcinoma Survivors | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A Machine Learning–Based Prediction Model for Dysphagia After Radiotherapy in Nasopharyngeal Carcinoma Survivors Yan He, Hui Zhao, Yongjiao Kang, Jiajie Xu, Yanling Wen, Jia Li, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9450763/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Dysphagia is a common long-term complication in patients with nasopharyngeal carcinoma following radiotherapy, significantly impairing nutritional status and quality of life. Therefore, early identification of high-risk individuals and timely intervention are of critical importance. However, there is currently a lack of effective risk prediction tools that integrate multidimensional symptom profiles with clinical characteristics. This study aims to develop a machine learning–based predictive model for dysphagia and to evaluate its predictive performance, thereby providing a reference for early identification and prevention strategies. Methods A total of 499 nasopharyngeal carcinoma survivors who attended a tertiary cancer hospital in Guangzhou between October 2023 and November 2025 were included in this study. Participants were randomly divided into a training set (n = 349) and a testing set (n = 150) at a ratio of 7:3. Data on demographic characteristics, clinical variables, and patient-reported outcomes were collected, including the M.D. Anderson Dysphagia Inventory (MDADI), Voice Handicap Index (VHI), M.D. Anderson Symptom Inventory–Head and Neck Module (MDASI-H&N), Generalized Anxiety Disorder-7 (GAD-7), and Patient Health Questionnaire-9 (PHQ-9). Predictor variables were selected using univariate analysis and least absolute shrinkage and selection operator (LASSO) regression. Three predictive models were subsequently developed: logistic regression (LR), backpropagation neural network (BPNN), and classification and regression tree (CART). Model discrimination was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. Calibration curves, Brier scores, and Shapley Additive Explanations (SHAP) were further applied to assess model performance and interpretability. Results The incidence of post-radiotherapy dysphagia among nasopharyngeal carcinoma survivors was 39.3%. A total of 14 predictive variables were ultimately selected. In the training set, both the LR and CART models achieved an AUC of 0.824, outperforming the BPNN model (AUC = 0.783). In the testing set, the BPNN model demonstrated the best performance (AUC = 0.776, accuracy = 0.720), exceeding that of the LR model (AUC = 0.666) and the CART model (AUC = 0.753). SHAP analysis indicated that depression, voice handicap index, gastrointestinal/oral symptom cluster, tumor stage, N stage, neck dysfunction, targeted therapy, and 3–5 years post-radiotherapy were the most influential contributors to the BPNN model. Conclusions The BPNN model demonstrated good stability and generalizability in predicting post-radiotherapy dysphagia in nasopharyngeal carcinoma survivors, and may facilitate early identification of high-risk patients and the implementation of targeted clinical interventions. Nasopharyngeal carcinoma Radiotherapy Dysphagia Risk prediction model Machine learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Nasopharyngeal carcinoma (NPC) is highly prevalent in southern China, with an incidence ranging from 25 to 30 per 100,000 population 1 . The adoption of comprehensive treatment strategies, primarily based on intensity-modulated radiation therapy (IMRT) combined with chemotherapy, has increased the 5-year survival rate of NPC patients to over 85% 2 . However, while high-dose radiotherapy effectively eradicates tumors, it inevitably causes damage to surrounding normal tissues, resulting in a range of late complications 3 . Among these, dysphagia is one of the most common and severe sequelae, with a reported incidence ranging from 13% to 93.5% 4 . Dysphagia not only leads to physical complications such as aspiration, malnutrition, and aspiration pneumonia, but also significantly impairs patients' psychological well-being and quality of life 5 . Notably, dysphagia is characterized by a delayed onset and progressive deterioration, with some survivors experiencing functional decline even 2–10 years after the completion of radiotherapy 6 . However, most existing studies have focused on swallowing function during or shortly after radiotherapy, with limited attention given to the dynamic changes and risk prediction of dysphagia among long-term survivors 7 . In clinical practice, early symptoms are often overlooked by patients, and issues such as underdiagnosis and suboptimal management remain prevalent 7 . Therefore, systematic assessment and risk prediction of swallowing function in NPC survivors after radiotherapy are of substantial clinical importance for delaying disease progression, reducing hospitalization, and lowering healthcare costs. Previous studies have primarily employed traditional statistical methods, such as LR and Cox regression, to identify risk factors for dysphagia 8 . However, clinical data are often characterized by nonlinearity, complex interactions among variables, and missing values, which limit the predictive accuracy of conventional approaches 9 . In recent years, machine learning, with its strong capabilities in data mining and self-learning, has demonstrated significant advantages in handling high-dimensional and complex medical data and in developing accurate predictive models 10 . For instance, Du et al. 11 applied machine learning techniques to predict the risk of breast cancer–related lymphedema in 670 patients, achieving promising results. Nevertheless, the application of machine learning in predicting post-radiotherapy dysphagia in NPC remains at an early stage and warrants further investigation and validation. In this study, we analyzed factors associated with dysphagia in NPC survivors and employed univariate analysis and least absolute shrinkage and selection operator (LASSO) regression for variable selection. Three predictive models—logistic regression (LR), backpropagation neural network (BPNN), and classification and regression tree (CART)—were constructed and compared to identify the optimal model. The aim was to facilitate early identification of high-risk NPC survivors following radiotherapy and to support the implementation of targeted interventions to reduce the incidence of dysphagia. Methods Research design This study employed a cross-sectional design. Sample and setting in research A convenience sampling method was used to recruit NPC survivors who attended a tertiary Grade A hospital in Guangdong Province between October 2023 and November 2025. The inclusion criteria were as follows: (1) histopathological confirmation of NPC based on imaging and endoscopic biopsy according to the 8th edition of the AJCC staging system 12 ; (2) age ≥ 18 years; (3) completion of radiotherapy; (4) provision of informed consent and voluntary participation; and (5) adequate cognitive and communication abilities. The exclusion criteria included: (1) severe cognitive impairment, psychiatric disorders, or other conditions affecting questionnaire completion and communication; and (2) critical illness or the presence of other severe organic diseases. For the development of the prediction models, a total of 21 candidate predictors were initially identified based on relevant literature and institutional data 13 . According to the events per variable principle, a minimum of 210 positive events (21 × 10) was required for model construction. Based on preliminary pilot data, the incidence of dysphagia among NPC survivors was approximately 57%. Therefore, the minimum required sample size was estimated to be 369 cases (210 ÷ 0.57). Considering a potential invalid response rate of 10%–20%, a total of 410–462 participants were deemed necessary. Model performance was evaluated using cross-validation, with the dataset randomly divided into a training set (70%) and a testing set (30%). Data Collection Face-to-face surveys were conducted by two specially trained nurses when NPC survivors returned for follow-up visits. Participants were informed of the study objectives and instructions for questionnaire completion, and written informed consent was obtained prior to administration. Standardized instructions were provided, and questionnaires were self-administered. For participants who had difficulty completing the questionnaire, trained investigators assisted by recording responses based on the participants' verbal reports with their consent. Each interview lasted approximately 15–30 minutes and was completed during waiting periods. Disease-related information was supplemented by reviewing the electronic medical record system. All questionnaires were collected and checked on site; those with more than 10% missing data or with patterned responses were excluded as invalid. Data were double-checked by two researchers before entry. A total of 542 questionnaires were distributed, of which 499 were valid, yielding an effective response rate of 92.06%. Ethical Considerations This study was approved by the Medical Ethics Committee of the authors' institution (approval number: SZR2022-179), and all participants provided informed consent prior to enrollment. Instruments Sociodemographic and Clinical Characteristics Based on a review of the literature and expert consultation, a self-designed questionnaire was developed by the researchers. It included sociodemographic variables (age, sex, place of residence, marital status, educational level, monthly household income per capita, return-to-work status, and adherence to swallowing training) and disease-related variables (treatment stage, AJCC stage, T/N/M classification, radiotherapy modality, receipt of induction chemotherapy, immunotherapy, and targeted therapy). M.D. Anderson Dysphagia Inventory (MDADI) The M.D. Anderson Dysphagia Inventory (MDADI), developed by Chen et al. 14 in 2001, consists of 20 items across four domains: global, emotional, functional, and physical. Each item is rated on a 5-point Likert scale ranging from "strongly agree" to "strongly disagree" (scored 1–5), with higher scores indicating better swallowing function and quality of life. In this study, a score > 69 was used to define the presence of dysphagia. The Cronbach's α coefficient of the scale is 0.82 15 . voice index (VHI) The voice index (VHI), proposed by Jacobson et al. 16 in 1997, assesses voice-related impairment across three domains: functional, physical, and emotional. It includes 30 items, each scored from 0 ("never") to 4 ("always"), yielding a total score ranging from 0 to 120. Higher scores indicate greater perceived voice impairment. The Cronbach's α coefficient is 0.86 17 . Neck Disability Index (NDI) The Neck Disability Index (NDI), developed by Vernon et al. 18 in 1991, consists of 10 items, each scored from 0 to 5, with a total score ranging from 0 to 50. Higher scores indicate more severe functional impairment. The Cronbach's α coefficient of the scale is > 0.81 19 . M.D. Anderson Symptom Inventory–Head and Neck Module (MDASI-H&N) The MDASI-H&N, developed by Rosenthal et al. 20 , is based on the M.D. Anderson Symptom Inventory and consists of two parts. The first part assesses the severity of 22 symptoms experienced in the past 24 hours, using an 11-point Likert scale ranging from 0 ("not present") to 10 ("as severe as you can imagine"), with higher scores indicating greater symptom severity. The second part evaluates the extent to which symptoms interfere with daily life and includes 6 items. The Chinese version was translated and validated by Han Yuan et al. 21 in 2010, with Cronbach's α coefficients of 0.877 and 0.835 for the core and module-specific items, respectively. Generalized Anxiety Disorder-7 (GAD-7) The Generalized Anxiety Disorder-7 (GAD-7), developed by Spitzer et al. 22 based on the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV), consists of 7 items rated on a 4-point Likert scale from "not at all" to "nearly every day," with total scores ranging from 0 to 21. Cut-off scores of 5, 10, and 15 represent mild, moderate, and severe anxiety, respectively. The Cronbach's α coefficient is 0.949 23 . Patient Health Questionnaire-9 (PHQ-9) The Patient Health Questionnaire-9 (PHQ-9) uses a 4-point Likert scale (0 = "not at all" to 3 = "nearly every day"), with total scores ranging from 0 to 27. Cut-off scores of 5, 10, and 15 indicate mild, moderate, and severe depression, respectively 24 . In this study, a PHQ-9 score ≥ 5 was used as the screening threshold for depression. The Cronbach's α coefficient of the scale is 0.917 25 . Treatment Regimens Radiotherapy Target volume delineation was performed in accordance with the International Commission on Radiation Units and Measurements (ICRU) Reports No. 50. The gross tumor volume (GTV) included the primary tumor and metastatic cervical lymph nodes 26 . The clinical target volume (CTV) was individually defined based on the extent of tumor infiltration. Planning target volumes (PTVs) were generated by adding a 3-mm margin to the CTV to account for setup uncertainties. The prescribed radiation doses were 70 Gy to the GTV, 66–70 Gy to involved lymph nodes, 60–62 Gy to the high-risk CTV, and 54–56 Gy to the low-risk CTV, delivered in 30–33 fractions. IMRT and helical tomotherapy (HT) plans were assigned to experienced medical physicists, and all treatment plans were reviewed and approved by senior radiation oncologists. Treatment was delivered using a synchronized integrated boost technique, with one fraction per day, five days per week, over 6–7 weeks. IMRT plans were generated using nine coplanar fields. Optimization was performed using the dose-volume optimizer algorithm in the Eclipse treatment planning system (TPS). The plans were delivered using a linear accelerator equipped with a dynamic multileaf collimator system (NOMOS Corporation, Sewickley, PA, USA) in a slice-by-slice arc rotation technique 27 . HT plans were generated using a tomotherapy planning system with 6-MV photon beams. The main optimization parameters included a field jaw width of 1.0 cm, a pitch of 0.287, and a modulation factor of 3.8. Dose calculations were performed using a convolution–superposition algorithm with a fine calculation grid of 0.273 cm × 0.273 cm × 0.3 cm 28 . Chemotherapy Induction chemotherapy was administered according to the patient's clinical condition, using platinum-based regimens at 21-day intervals. Common induction chemotherapy (IC) protocols included docetaxel–cisplatin–5-fluorouracil (TPF), docetaxel–cisplatin (TP), cisplatin–5-fluorouracil (PF), and gemcitabine–cisplatin (GP). Concurrent chemotherapy during radiotherapy consisted of cisplatin at a dose of 80 or 100 mg/m² for 2–3 cycles 29 . Data analysis Statistical analyses were performed using SPSS version 27.0 (IBM Corp., Armonk, NY, USA), and machine learning–based model development and validation were conducted using Python. For descriptive analysis, continuous variables with a non-normal distribution were presented as median (interquartile range, M [P25, P75]) and compared using the Mann–Whitney U test. Categorical variables were expressed as frequencies (percentages) and compared using the chi-square test or Fisher's exact test when expected cell counts were < 5. Normally distributed continuous variables were reported as mean ± standard deviation (x̄ ± s) and compared using independent-samples t-tests. Exploratory factor analysis was performed using maximum variance orthogonal rotation to extract symptom factors. Factors were retained according to the following criteria: eigenvalues ≥ 1, inclusion of at least two symptoms per factor, and factor loadings ≥ 0.4. When a symptom loaded ≥ 0.4 on multiple factors, it was assigned to the factor with the highest loading 30 . For predictive modeling, Python was used to construct three models: LR, BPNN, and CART. Variables identified through univariate LR were further selected using LASSO regression. All features were standardized using Z-score normalization. The optimal regularization parameter (α) was determined via 10-fold cross-validation, and variables with non-zero coefficients were retained as final predictors. The selected features were subsequently used to build multivariable LR, BPNN, and CART models. The dataset was randomly split into a training set (n = 349) and a validation set (n = 150) at a ratio of 7:3 31 . (1) The CART model was constructed using the Gini index as the splitting criterion. To prevent overfitting, the maximum tree depth was restricted to 4, and class-weight balancing was applied. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, calibration curves, and Brier scores in both the training and validation sets. (2) The BPNN model was developed using standardized continuous variables. The network consisted of two hidden layers (30 and 15 neurons, respectively), with ReLU activation functions and the Adam optimizer. L2 regularization and early stopping were applied to reduce overfitting. Feature importance was assessed using permutation importance. Model performance was evaluated using ROC curves (AUC), calibration curves, and Brier scores. (3) The multivariable LR model was used to identify independent risk factors for dysphagia, with odds ratios (ORs) and 95% confidence intervals (CIs) reported. Model discrimination was assessed using ROC curves (AUC, sensitivity, and specificity), while calibration was evaluated using calibration curves and Brier scores. Finally, the optimal predictive model for post-radiotherapy dysphagia in NPC survivors was selected based on comprehensive performance evaluation in both the training and validation sets. Results Participant Characteristics A total of 499 participants were included in this study, among whom 196 survivors developed dysphagia, yielding an overall incidence of 39.3%. In the training set, 143 of 349 survivors experienced dysphagia, corresponding to an incidence of 41.0%. In the validation set, 53 of 150 patients developed dysphagia, with an incidence of 35.3%. There was no statistically significant difference in the incidence of dysphagia between the two groups ( χ² = 1.400, P > 0.05). The mean age of the participants was 45.55 ± 11.05 years. The proportion of males (63.5%) to females (36.5%) was approximately 2:1. Patients with stage III/IV disease accounted for 80.1% of the sample. The majority of survivors were married, comprising 85.8% of the study population. Detailed sociodemographic and clinical characteristics are presented in Table 1. Table 1 General Characteristics of NPC Survivors After Radiotherapy ( N =499) Variables Training Set (n=349) Validation Set (n=150) χ²/Z P Variables Sex 0.445 0.505 Male 225 (64.5%) 92 (61.3%) Female 124 (35.5%) 58 (38.7%) Age, years 45 (17) 43 (13) -1.398 0.162 Return to work 0.011 0.917 Yes 153 (43.8%) 65 (43.3%) No 196 (56.2%) 85 (56.7%) Residence 0.002 0.965 Urban 145 (41.5%) 62 (41.3%) Rural 204 (58.5%) 88 (58.7%) Marital status 0.009 0.924 Married 50 (14.3%) 21 (14.0%) Unmarried/divorced/widowed 299 (85.7%) 129 (86.0%) Education level 1.619 0.445 Middle school or below 136 (39.0%) 51 (34.0%) High school/vocational 73 (20.9%) 38 (25.3%) College or above 140 (40.1%) 61 (40.7%) Monthly income (CNY) 6.052 0.109 10000 46 (13.2%) 12 (8.0%) Adherence to swallowing exercises 3.241 0.072 None/partial 315 (90.3%) 127 (84.7%) Full 34 (9.7%) 23 (15.3%) Treatment stage 8.100 0.044 5 years 40 (11.5%) 23 (15.3%) AJCC stage 1.887 0.596 I–II 50 (14.3%) 15 (10.0%) III 167 (47.9%) 78 (52.0%) IV 108 (30.9%) 47 (31.3%) Unknown 24 (6.9%) 10 (6.7%) T stage 0.794 0.851 T1–T2 43 (12.3%) 21 (14.0%) T3 225 (64.5%) 91 (60.7%) T4 73 (20.9%) 35 (23.3%) Unknown 8 (2.3%) 3 (2.0%) N stage 3.718 0.446 N0 28 (8.0%) 12 (8.0%) N1 119 (34.1%) 40 (26.7%) N2 104 (29.8%) 56 (37.3%) N3 90 (25.8%) 39 (26.0%) Unknown 8 (2.3%) 3 (2.0%) M stage 0.468 0.791 M0 327 (93.7%) 139 (92.7%) M1 14 (4.0%) 8 (5.3%) Unknown 8 (2.3%) 3 (2.0%) Radiotherapy technique 0.393 0.531 IMRT 320 (91.7%) 140 (93.3%) TOMO 29 (8.3%) 10 (6.7%) Induction chemotherapy 1.335 0.248 Yes 294 (84.2%) 120 (80.0%) No 55 (15.8%) 30 (20.0%) Immunotherapy 0.774 0.379 Yes 145 (41.5%) 56 (37.3%) No 204 (58.5%) 94 (62.7%) Targeted therapy 2.284 0.131 Yes 97 (27.8%) 32 (21.3%) No 252 (72.2%) 118 (78.7%) Dysphagia 1.400 0.237 Yes 143 (41.0%) 53 (35.3%) No 206 (59.0%) 97 (64.7%) Voice index 6.00 (12) 5.50 (12) -0.236 0.814 Depression 4.00 (8) 4.00 (7) -0.048 0.961 Anxiety 2.00 (7) 2.00 (7) -0.140 0.889 Gastrointestinal/oral symptoms 2.45 (3.09) 2.09 (3.02) -0.483 0.629 Emotional/fatigue symptoms 2.14 (2.93) 2.14 (2.90) -0.235 0.814 Function/quality of life symptoms 1.33 (3.09) 1.33 (3.00) -0.673 0.501 Cervical dysfunction level 6.00 (7) 6.00 (7) -0.041 0.967 Univariate Analysis of Dysphagia Prior to univariate regression analysis, symptom clusters were categorized into three groups based on exploratory factor analysis loadings: the gastrointestinal/oral symptom cluster, emotional/fatigue symptom cluster, and function/quality of life cluster (see Supplementary Table S1 for details). The results of the univariate analysis indicated that multiple variables were significantly associated with the occurrence of dysphagia in the training set ( p < 0.05). Among these, sociodemographic factors included return-to-work status, residence, educational level, monthly income, and age. Clinical factors comprised AJCC stage, N stage, and receipt of immunotherapy and targeted therapy. Psychological factors included levels of depression and anxiety. Symptom- and function-related variables included voice index, gastrointestinal/oral symptom cluster, emotional/fatigue symptom cluster, function/quality of life cluster, and cervical dysfunction level. Detailed results are presented in Supplementary Table S2. LASSO Regression Analysis of Dysphagia LASSO regression was performed to further select predictive variables for post-radiotherapy dysphagia in NPC survivors. Variables identified as statistically significant in the univariate analysis were included in the LASSO model. As the regularization parameter ( λ ) increased, the mean squared error (MSE) initially decreased and then increased. At the optimal value of λ , the regression coefficients of residence and anxiety were shrunk to zero and thus excluded from the final model. Ultimately, 14 variables with relatively stable non-zero coefficients were retained, suggesting that they may be important predictors of dysphagia and were subsequently included in the risk prediction models. Detailed results are presented in Figures 1 and 2. Figure 1 LASSO Coefficient Profiles Figure 2 Cross-Validation Curve for LASSO Regression Model Development and Evaluation In this study, three predictive models—CART, BPNN, and LR—were developed to estimate the risk of post-radiotherapy dysphagia in patients with NPC. Model performance was evaluated in both the training and validation sets, with discrimination assessed using ROC curves and the AUC. In the training set, the AUC values for the CART, BPNN, and LR models were 0.824, 0.783, and 0.824, respectively. In the validation set, the corresponding AUC values were 0.666, 0.776, and 0.753, indicating overall good predictive performance (see Figure 3 and Supplementary Figures S1–S3). Comparative analysis showed that CART and LR achieved the highest discrimination in the training set (AUC = 0.824). However, in the validation set, the BPNN model demonstrated superior robustness, with the highest AUC (0.776) and accuracy (0.720) among the three models. In terms of specificity, both the BPNN and LR models exhibited high performance in the validation set (0.845 and 0.825, respectively), indicating strong capability in correctly identifying patients at risk of dysphagia. Overall, the BPNN model demonstrated the best comprehensive performance and may serve as a promising tool for clinical application. Detailed results are presented in Table 3. Table 4 Comparison of predictive performance among the three models Training Set Validation Set Metrics CART BPNN LR CART BPNN LR AUC 0.824 0.783 0.824 0.666 0.776 0.753 Accuracy 0.777 0.745 0.777 0.653 0.720 0.700 Sensitivity 0.748 0.587 0.664 0.623 0.528 0.472 Specificity 0.796 0.854 0.854 0.670 0.845 0.825 Brier Score - 0.181 0.159 - 0.180 0.186 Notes: AUC, area under the curve; CI, confidence interval. DeLong test used for comparisons. Importance Ranking of Predictors in the Neural Network Model The results (Figure 4) indicated that, as measured by the decrease in AUC, depression was the most important predictor (0.0650), substantially exceeding all other variables. This was followed by the voice index (0.0403) and the gastrointestinal/oral symptom cluster (0.0209). In addition, N stage (N2: 0.0128; N3: 0.0105), Cervical dysfunction level (0.0115), and targeted therapy (0.0077) also demonstrated notable predictive value. In contrast, sociodemographic variables such as age and income showed relatively low importance, with some even exhibiting negative contributions, suggesting limited impact on improving model performance. The SHAP analysis results in the testing set were generally consistent with these findings, highlighting the dominant role of symptom burden and psychological factors in the model. Furthermore, the SHAP summary plot illustrated the directional effects of predictors on model output: higher levels of depression, greater symptom burden, and more severe cervical dysfunction were all associated with an increased risk of dysphagia. By comparison, certain sociodemographic variables (e.g., income level and return-to-work status) showed weaker and more variable effects across individuals. Overall, the predictive performance of the model was primarily driven by psychological status, symptom burden, and functional impairment, whereas traditional sociodemographic factors contributed relatively little. These findings suggest that post-radiotherapy dysphagia in NPC survivors is characterized by complex, multidimensional interactions. Detailed results are presented in Supplementary Figure S4. Discussion The high incidence of dysphagia among NPC survivors highlights the need to establish a comprehensive, longitudinal monitoring system. In this study, the incidence of post-radiotherapy dysphagia among NPC survivors was 39.3%, which is generally consistent with the 42.8% reported by Li et al. 7 . Currently, IMRT has replaced conventional radiotherapy as the standard treatment modality for NPC. Through precise spatial dose modulation, IMRT effectively reduces radiation exposure to surrounding normal tissues, thereby lowering the incidence of dysphagia compared with conventional techniques 3 . However, due to the unique anatomical location of the nasopharynx—adjacent to swallowing-related muscles and cranial nerves—radiation-induced damage remains difficult to completely avoid. From a pathophysiological perspective, radiation-induced injury to cranial nerves may result in paralysis and sensory impairment of the pharyngeal and laryngeal muscles, leading to delayed swallowing reflexes and aspiration. In addition, direct damage to the swallowing musculature can cause fibrosis and atrophy, ultimately impairing swallowing efficiency 26 . Notably, dysphagia is characterized by a delayed onset and progressive worsening, with some patients developing symptoms months or even years after the completion of radiotherapy 6 . McDowell et al. 32 reported in a long-term follow-up study of NPC survivors treated with IMRT that a considerable proportion of patients continued to experience swallowing and chewing difficulties even 7.5 years post-treatment, which were closely associated with emotional distress, including depression and anxiety. Therefore, in clinical practice, it is essential to establish a dynamic, longitudinal assessment framework beginning at the initiation of radiotherapy. Regular monitoring of swallowing function and structured follow-up should be implemented to facilitate early identification of high-risk individuals and enable timely, individualized interventions, ultimately reducing the incidence of dysphagia and improving long-term quality of life. Comparison of the predictive performance and clinical applicability of different machine learning models This study, based on clinical data from 499 NPC patients after radiotherapy, systematically compared the predictive performance of three machine learning models (CART, BPNN, and LR) for post-radiotherapy dysphagia. The results demonstrated that the BPNN model exhibited the best robustness in the validation set, with a high specificity of 0.845, indicating its strong ability to correctly identify patients without dysphagia. In contrast, the LR and CART models showed relatively higher discrimination in the training set, suggesting potential limitations in generalizability. Overall, the BPNN model demonstrated superior clinical applicability and potential value for identifying high-risk populations. Srinivasan et al. 33 , in a recent review, reported that machine learning models, by integrating multimodal data, exhibit superior discriminative ability compared with traditional models in predicting treatment-related toxicities such as dysphagia and voice impairment in head and neck cancer patients. The BPNN model can automatically capture nonlinear relationships and complex interactions among variables, and continuously optimize model weights through the backpropagation algorithm, thereby improving predictive accuracy. This finding is consistent with the results reported by Nicol et al. 34 . A large-scale study involving 1,685 head and neck cancer patients 35 also demonstrated that mixed-effect models incorporating dosimetric factors achieved stable discriminative performance in predicting dysphagia (AUC = 0.77–0.84). However, the CART model showed evidence of overfitting in the validation set (AUC: 0.824 in the training set vs. 0.666 in the validation set), indicating limited generalizability. In contrast, the BPNN model demonstrated more stable predictive performance and may be better suited for clinical decision-support systems, although its reliability should still be validated in specific clinical contexts. Multidimensional determinants of dysphagia risk in NPC survivors This study identified depression as the most important predictor in the BPNN-based feature importance analysis, highlighting the critical role of psychological factors in swallowing rehabilitation. Wang et al. 36 reported that 44.2% of 773 post-radiotherapy NPC survivors experienced mild to moderate depressive symptoms. Similarly, McDowell et al. 32 found that among NPC survivors treated with IMRT, the prevalence of depression and anxiety was 25% and 37%, respectively, both of which were significantly negatively associated with quality of life. Dysphagia-related eating difficulties, nutritional risk, and social withdrawal may directly lead to or exacerbate depressive symptoms. In turn, impaired swallowing function may cause patients to avoid public eating, alter dietary patterns, and experience disruption in daily life routines, with some individuals adopting maladaptive coping strategies that further reduce quality of life 5 . Moreover, depression may amplify the subjective perception of dysphagia, pain, and fatigue, and may also influence physiological recovery through neuroendocrine and inflammatory pathways, thereby creating a bidirectional "symptom–psychology" vicious cycle that further worsens swallowing dysfunction 7 . Therefore, rehabilitation strategies focusing solely on swallowing function may be insufficient, and integrated approaches combining swallowing rehabilitation with psychological intervention are warranted in clinical practice. The voice index ranked as the second most important predictor, suggesting a close anatomical and functional association between voice and swallowing dysfunction, which is consistent with the findings of Wentzel et al. 37 . Swallowing and phonatory functions share common anatomical structures, including the tongue, soft palate, pharynx, and larynx, and damage to these structures or their neural control may simultaneously impair both functions 38 . Portas et al. 39 demonstrated in patients after laryngeal cancer surgery that voice index can indicate the presence of swallowing impairment. Therefore, systematic collection of patient-reported voice outcomes (e.g., VHI) during clinical follow-up may not only facilitate voice rehabilitation management but also provide important clues for early identification of dysphagia risk. This study further found that the gastrointestinal/oral symptom cluster was highly influential in the model, indicating that dysphagia is not an isolated event but is closely associated with multiple co-occurring symptoms. Previous studies have confirmed that symptoms in NPC survivors tend to cluster, with xerostomia, pain, taste alteration, and dysphagia interacting synergistically to impair eating behavior and quality of life 40 . A cross-sectional study by Tam et al. 41 involving 211 survivors further supported this view, showing that nearly all patients experienced late radiation-related toxicities, and more than three-quarters reported at least four moderate-to-severe symptoms. Dysphagia and xerostomia were among the top five most common symptoms, and symptom burden was significantly positively correlated with unresolved symptom distress, with no improvement over time. From a pathophysiological perspective, radiation-induced damage to salivary glands and pharyngeal muscles leads to impaired bolus formation and pharyngeal residue, thereby increasing the risk of dysphagia 42 . These findings suggest that post-radiotherapy dysphagia management in NPC survivors should move beyond a single-symptom approach toward an integrated strategy combining symptom cluster monitoring and functional rehabilitation. In addition, cervical dysfunction and N stage remained important predictors, suggesting that radiation-induced structural damage and tumor burden play fundamental roles in dysphagia development. Post-radiotherapy fibrosis and atrophy of cervical muscles lead to restricted neck mobility, weakened laryngeal musculature, impaired pharyngoesophageal movement, and reduced neural reflex function, ultimately resulting in poor coordination of swallowing and mastication 43 . Although IMRT has improved survival outcomes, more intensive treatment regimens are also associated with a higher incidence of dysphagia 44 . Moreover, higher N stage generally indicates a wider irradiation field and more extensive tissue involvement. Reduced regenerative capacity, reactive exudation, tissue sclerosis, and fibrosis within the irradiated field may lead to trismus and restricted soft tissue mobility, further increasing dysphagia risk 45 , 46 . Targeted therapy, while enhancing anti-tumor effects through inhibition of angiogenesis and immune-mediated tumor suppression 47 , may also increase the incidence of skin reactions and oral mucositis, thereby exacerbating injury to the orofacial and cervical musculature. Finally, SHAP analysis based on the validation set further elucidated both the magnitude and directionality of variable contributions. Depression, symptom cluster burden, and cervical dysfunction consistently exerted positive effects on dysphagia risk, indicating that higher levels of these factors were associated with increased risk across individuals. In contrast, certain socioeconomic variables (e.g., income level and return-to-work status) showed heterogeneous effects across individuals, suggesting context-dependent influences. Compared with traditional statistical methods, SHAP enables quantification of marginal contributions at the individual level and reveals potential nonlinear relationships and interaction effects, thereby enhancing model interpretability and clinical applicability 48 . Limitations Despite exploring the determinants of dysphagia from a multidimensional perspective, several limitations should be acknowledged. First, future studies should conduct external validation using multicenter, large-scale datasets to enhance the generalizability and clinical applicability of the model. Second, incorporating imaging features, biological markers, and objective functional assessments (e.g., videofluoroscopic swallowing study or electromyography) may help develop more refined predictive models and further improve predictive accuracy. Third, complex interactions and potential causal relationships among variables may exist. Future studies employing causal inference methods or longitudinal designs are needed to elucidate the dynamic evolution of psychological factors, symptom clusters, and functional impairment in the development of dysphagia. Finally, based on the key risk factors identified in this study, interventional research should be conducted to evaluate the effectiveness of integrated psychological support, symptom management, and functional rehabilitation strategies in improving swallowing outcomes, thereby enhancing the long-term quality of life of NPC survivors. Conclusion Based on clinical data from 499 patients with NPC after radiotherapy, this study developed dysphagia risk prediction models using three machine learning algorithms, including LR, BPNN and CART. The results showed that the incidence of post-radiotherapy dysphagia was 39.3%. The BPNN model demonstrated the best predictive robustness and relatively high specificity in the validation set, suggesting its potential value for clinical risk stratification and screening. SHAP analysis further indicated that depression, emotional/fatigue symptom clusters, and cervical dysfunction were the main predictive factors. Based on these findings, the management of post-radiotherapy dysphagia in NPC survivors should move beyond a sole focus on structural injury and shift toward a multidimensional risk stratification strategy that integrates psychological status, symptom clusters, and cervical functional training, in order to achieve early identification, precise intervention, and improved long-term quality of life. Declarations Conflict of interest The authors made no disclosures. Funding Sun Yat-sen University Cancer Center Yue-Qin Nursing Research and Innovation Fund, YQ2025001-A, YQ2025007-B. Author Contribution Yan He, Hui Zhao, Yongjiao Kang contributed equally to this work and should be considered co-first authors. Yan He, Hui Zhao, Yongjiao Kang were responsible for the study conception, design, and manuscript drafting. Jiajie Xu, Yanling Wen contributed to data collection and management. Jia Li, Yuying Fan and Wen Hu provided critical revisions and guidance on study methodology, statistical analysis, and interpretation of results. All authors read and approved the final manuscript. Jia Li, Yuying Fan, and Wen Hu are joint corresponding authors. Acknowledgements None. Data availability statements The datasets used and/or analysed during the current study available from the corresponding author on reasonable request. References Chen Y-P, Chan ATC, Le Q-T, Blanchard P, Sun Y, Ma J. Nasopharyngeal carcinoma. Lancet. 2019;394:64–80. Wei X, Chen B, Wang Z, Zhao P, Duan X. Nasopharyngeal cancer risk assessment by country or region worldwide from 1990 to 2019. BMC Public Health. 2024;24:1931. Yao J-J, Zhang F, Gao T-S, Zhang W-J, Lawrence WR, Zhu B-T, et al. Survival impact of radiotherapy interruption in nasopharyngeal carcinoma in the intensity-modulated radiotherapy era: A big-data intelligence platform-based analysis. Radiother Oncol. 2019;132:178–87. Fong R, Ward EC, Rumbach AF. Dysphagia after chemo-radiation for nasopharyngeal cancer: A scoping review. World J Otorhinolaryngol Head Neck Surg. 2020;6:10–24. McCarty EB, Chao TN. Dysphagia and Swallowing Disorders. Med Clin North Am. 2021;105:939–54. Suiter DM, Gosa MM. Assessing and Treating Dysphagia: A Lifespan Perspective . Thieme: Stuttgart; 2019. Li H, Li L, Huang X, Li Y, Zou T, Zhuo X, et al. Radiotherapy-induced dysphagia and its impact on quality of life in patients with nasopharyngeal carcinoma. Strahlenther Onkol. 2019;195:457–67. Langendijk JA, Doornaert P, Rietveld DHF, Verdonck-de Leeuw IM, Leemans CR, Slotman BJ. A predictive model for swallowing dysfunction after curative radiotherapy in head and neck cancer. Radiother Oncol. 2009;90:189–95. Atias D, Ashri S, Goldbourt U, Benyamini Y, Gilad-Bachrach R, Hasin T, et al. Machine learning in epidemiology: An introduction, comparison with traditional methods, and a case study of predicting extreme longevity. Ann Epidemiol. 2025;110:23–33. de Vette SPM, Neh H, van der Hoek L, MacRae DC, Chu H, Gawryszuk A, et al. Deep learning NTCP model for late dysphagia after radiotherapy for head and neck cancer patients based on 3D dose, CT and segmentations. Radiother Oncol. 2025;213:111169. Du J, Yang J, Yang Q, Zhang X, Yuan L, Fu B. Comparison of machine learning models to predict the risk of breast cancer-related lymphedema among breast cancer survivors: a cross-sectional study in China. Front Oncol. 2024;14:1334082. Pan JJ, Ng WT, Zong JF, Chan LLK, O’Sullivan B, Lin SJ, et al. Proposal for the 8th edition of the AJCC/UICC staging system for nasopharyngeal cancer in the era of intensity-modulated radiotherapy. Cancer. 2016;122:546–58. Riley RD, Ensor J, Snell KIE, Harrell FE, Martin GP, Reitsma JB, et al. Calculating the sample size required for developing a clinical prediction model. BMJ. 2020;368:m441. Chen AY, Frankowski R, Bishop-Leone J, Hebert T, Leyk S, Lewin J, et al. The development and validation of a dysphagia-specific quality-of-life questionnaire for patients with head and neck cancer: the M. D. Anderson dysphagia inventory. Arch Otolaryngol Head Neck Surg. 2001;127:870–6. Yee K, Wong SM, Teo I, Loy J, Roche E, Tan YP, et al. Validity and reliability of the MD Anderson dysphagia inventory in English and Chinese in head and neck cancer patients. Asia Pac J Clin Oncol. 2020;16:372–9. Johnson A, Jacobson B, Grywalski C, Silbergleit A, Jacobson G, Benninger M. The Voice Handicap Index (VHI): Development and Validation. Am J Speech-Language Pathol. 1997;6:66–70. Xu W, Li H-Y, Hu R, Hu H-Y, Hou L-Z, Zhang L, et al. Analysis of reliability and validity of the Chinese version of voice handicap index (VHI). Chin J Otorhinolaryngol Head Neck Surg. 2008;43:670–5. Vernon H, Mior S. The Neck Disability Index: a study of reliability and validity. J Manipulative Physiol Ther. 1991;14:409–15. Saltychev M, Pylkäs K, Karklins A, Juhola J. Psychometric properties of neck disability index - a systematic review and meta-analysis. Disabil Rehabil. 2024;46:5415–31. Rosenthal DI, Mendoza TR, Chambers MS, Burkett VS, Garden AS, Hessell AC, et al. The M. D. Anderson symptom inventory-head and neck module, a patient-reported outcome instrument, accurately predicts the severity of radiation-induced mucositis. Int J Radiat Oncol Biol Phys. 2008;72:1355–61. Han Y, Zhang MF, Zhang JE. Symptom distress during and after radiation therapy in patients with nasopharyngeal carcinoma[J]. Chin J Nurs. 2010;45(07):626–8. Spitzer RL, Kroenke K, Williams JBW, Löwe B. A brief measure for assessing generalized anxiety disorder: the GAD-7. Arch Intern Med. 2006;166:1092–7. Gong Y, Zhou H, Zhang Y, Zhu X, Wang X, Shen B, et al. Validation of the 7-item Generalized Anxiety Disorder scale (GAD-7) as a screening tool for anxiety among pregnant Chinese women. J Affect Disord. 2021;282:98–103. Spitzer RL, Kroenke K, Williams JB. Validation and utility of a self-report version of PRIME-MD: the PHQ primary care study. Primary Care Evaluation of Mental Disorders. Patient Health Questionnaire JAMA. 1999;282:1737–44. Wang W, Bian Q, Zhao Y, Li X, Wang W, Du J, et al. Reliability and validity of the Chinese version of the Patient Health Questionnaire (PHQ-9) in the general population. Gen Hosp Psychiatry. 2014;36:539–44. Zhang Q-F, Zhou K-B, Chen Z-J, Long Y-B. Pathogenesis and progress of rehabilitation treatment of common complications after radiotherapy for nasopharyngeal carcinoma. Chin J Otorhinolaryngol - Skull Base Surg. 2023;29:102–7. Liu Q, Yao W-Y, Xu P-X, Fang Y-W, Lin Z-Y, Li Z-T, et al. Comparison of radiation-induced temporal lobe injury after helical tomotherapy versus conventional intensity-modulated radiotherapy for nasopharyngeal carcinoma: a propensity score-matched cohort study. Radiat Oncol. 2026;21:23. Li W-F, Li Y-Q, Chen L, Zhang Y, Guo R, Zhang F, et al. Propensity-matched analysis of three different chemotherapy sequences in patients with locoregionally advanced nasopharyngeal carcinoma treated using intensity-modulated radiotherapy. BMC Cancer. 2015;15:810. Lee N, Harris J, Garden AS, Straube W, Glisson B, Xia P, et al. Intensity-modulated radiation therapy with or without chemotherapy for nasopharyngeal carcinoma: radiation therapy oncology group phase II trial 0225. J Clin Oncol. 2009;27:3684–90. Beam AL, Kohane IS. Big Data and Machine Learning in Health Care. JAMA. 2018;319:1317–8. Hammer MJ, Harris C, Pozzar R, Snowberg K, Paul SM, Cooper BA, et al. Symptom Clusters and Symptom Burden in Cancer Survivors. Cancer Med. 2026;15:e71653. McDowell LJ, Rock K, Xu W, Chan B, Waldron J, Lu L, et al. Long-Term Late Toxicity, Quality of Life, and Emotional Distress in Patients With Nasopharyngeal Carcinoma Treated With Intensity Modulated Radiation Therapy. Int J Radiat Oncol Biol Phys. 2018;102:340–52. Srinivasan Y, Liu A, Rameau A. Machine learning in the evaluation of voice and swallowing in the head and neck cancer patient. Curr Opin Otolaryngol Head Neck Surg. 2024;32:105–12. Nicol AJ. Multi-omic prediction of severe acute treatment-induced oral mucositis and dysphagia in nasopharyngeal carcinoma patients. 2025 https://theses.lib.polyu.edu.hk/handle/200/13701 . Accessed 14 April 2026. Wordenskjold Stougaard S, Zukauskaite R, Röttger R, Laugaard Lorenzen E, Lukas Konrad M, Long Krogh S, et al. Impact of GTV-CTV margin and other predictors on radiation-induced dysphagia in head and neck cancer patients from DAHANCA group. Acta Oncol. 2025;64:1253–61. Wang S, Lai L, Deng Y, Zhang Y, Cai Y, Huang W, et al. The mediating role of resilience in the relationships between posttraumatic growth and depression as well as anxiety among survivors of nasopharyngeal carcinoma: a cross-sectional study. BMC Psychol. 2025;13:768. Wentzel A, Mohamed ASR, Naser MA, van Dijk LV, Hutcheson K, Moreno AM, et al. Multi-organ spatial stratification of 3-D dose distributions improves risk prediction of long-term self-reported severe symptoms in oropharyngeal cancer patients receiving radiotherapy: development of a pre-treatment decision support tool. Front Oncol. 2023;13:1210087. Dos Santos KW, da Cunha Rodrigues E, Rech RS, da Ros Wendland EM, Neves M, Hugo FN, et al. Using Voice Change as an Indicator of Dysphagia: A Systematic Review. Dysphagia. 2022;37:736–48. Portas JG, Queija D dos, Arine S, Ferreira LP, Dedivitis AS, Lehn RA. Voice and swallowing disorders: functional results and quality of life following supracricoid laryngectomy with cricohyoidoepiglottopexy. Ear Nose Throat J. 2009;88:E23–30. Li J, He Y, Xiao M-F, Huang X-T, Fan Y-Y, Lv X, Qin H-Y. Network analysis of symptom clusters and Quality of Life in Nasopharyngeal Carcinoma survivors. Chin Nurs Manage. 2025;25:1472–6. Tam VCW, Ching JCF, Yip SST, Kwong VHY, Chan CPL, Wong KCW, et al. Examining patient-reported late toxicity and its association with quality of life and unmet need for symptom management among nasopharyngeal cancer survivors: a cross-sectional survey. Front Oncol. 2024;14:1378973. Gómez Á, García-Chabur MA, Peñaranda D, Gómez-Mendoza A, Forero JC. Chemotherapy/Radiotherapy-Induced Dysphagia in Head and Neck Tumors: A Challenge for Otolaryngologists in Low- to Middle-Income Countries. Dysphagia. 2025;40:515–27. Gelblum D, Wolden S, Schupak K, Lee N. Neck Spasms as a Late Effect of Intensity Modulated Radiation Therapy (IMRT) for Head and Neck Cancer. Int J Radiat Oncol Biol Phys. 2011;81:S550. You R, Hua Y-J, Liu Y-P, Yang Q, Zhang Y-N, Li J-B, et al. Concurrent Chemoradiotherapy with or without Anti-EGFR-Targeted Treatment for Stage II-IVb Nasopharyngeal Carcinoma: Retrospective Analysis with a Large Cohort and Long Follow-up. Theranostics. 2017;7:2314–24. Carmignani I, Locatello LG, Desideri I, Bonomo P, Olmetto E, Livi L, et al. Analysis of dysphagia in advanced-stage head-and-neck cancer patients: impact on quality of life and development of a preventive swallowing treatment. Eur Arch Otorhinolaryngol. 2018;275:2159–67. Zheng Y, Han F, Xiao W, Xiang Y, Lu L, Deng X, et al. Analysis of late toxicity in nasopharyngeal carcinoma patients treated with intensity modulated radiation therapy. Radiat Oncol. 2015;10:17. Bonner JA, Harari PM, Giralt J, Azarnia N, Shin DM, Cohen RB, et al. Radiotherapy plus cetuximab for squamous-cell carcinoma of the head and neck. N Engl J Med. 2006;354:567–78. Linardatos P, Papastefanopoulos V, Kotsiantis S, Explainable AI. A Review of Machine Learning Interpretability Methods. Entropy (Basel). 2020;23:18. Additional Declarations No competing interests reported. Supplementary Files Supplementarymaterials.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9450763","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":630654667,"identity":"b9c17525-55e3-4ddc-844c-14f24a1dc1a7","order_by":0,"name":"Yan He","email":"","orcid":"","institution":"Sun Yat-sen University Cancer Center","correspondingAuthor":false,"prefix":"","firstName":"Yan","middleName":"","lastName":"He","suffix":""},{"id":630654668,"identity":"a6be7949-a9b8-4767-94a3-a719b52aed8f","order_by":1,"name":"Hui Zhao","email":"","orcid":"","institution":"Sun Yat-sen University Cancer Center","correspondingAuthor":false,"prefix":"","firstName":"Hui","middleName":"","lastName":"Zhao","suffix":""},{"id":630654670,"identity":"b20d5cc0-1810-4598-a155-97f45d253dad","order_by":2,"name":"Yongjiao Kang","email":"","orcid":"","institution":"Sun Yat-sen University Cancer Center","correspondingAuthor":false,"prefix":"","firstName":"Yongjiao","middleName":"","lastName":"Kang","suffix":""},{"id":630654671,"identity":"69a6bfca-7d8e-43dd-9b99-0c000f0ad4dc","order_by":3,"name":"Jiajie Xu","email":"","orcid":"","institution":"Sun Yat-sen University Cancer Center","correspondingAuthor":false,"prefix":"","firstName":"Jiajie","middleName":"","lastName":"Xu","suffix":""},{"id":630654672,"identity":"66dd73a0-c782-4d58-af6f-12b89fa26f09","order_by":4,"name":"Yanling Wen","email":"","orcid":"","institution":"Sun Yat-sen University Cancer Center","correspondingAuthor":false,"prefix":"","firstName":"Yanling","middleName":"","lastName":"Wen","suffix":""},{"id":630654673,"identity":"634f67f1-a966-423d-b39e-3d9d5f20e172","order_by":5,"name":"Jia Li","email":"","orcid":"","institution":"Sun Yat-sen University Cancer Center","correspondingAuthor":false,"prefix":"","firstName":"Jia","middleName":"","lastName":"Li","suffix":""},{"id":630654674,"identity":"3c6291fb-c6ff-4dab-ae00-43b45e67785b","order_by":6,"name":"Yuying Fan","email":"","orcid":"","institution":"Sun Yat-sen University Cancer Center","correspondingAuthor":false,"prefix":"","firstName":"Yuying","middleName":"","lastName":"Fan","suffix":""},{"id":630654675,"identity":"783dd341-26e7-4eeb-a3f5-6a75fd3ad27e","order_by":7,"name":"Wen Hu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAw0lEQVRIiWNgGAWjYBACAzBZQbqWMyRrYWwjRYu5RI7h48J5dYnzZyQ//MBQcc+ugf3sAbxaLGfkGBvP3HY4ccONNGMJhjPFyQ08eQn4HXYjd5s077YDiRskctiALkxIZpDgMSCkZftv3jkgh5GgZRszbwNzYsMNiBY7wlrOvP8szXPssPGGM8+MJRLOJCSw8eQQ0HI8LfEzT02d7Px2YIh9qEiw52c/g18LDDg2gMgEBoZEouPIHoMxCkbBKBgFowAGACLdQqYYRSmTAAAAAElFTkSuQmCC","orcid":"","institution":"Sun Yat-sen University Cancer Center","correspondingAuthor":true,"prefix":"","firstName":"Wen","middleName":"","lastName":"Hu","suffix":""}],"badges":[],"createdAt":"2026-04-17 15:23:32","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9450763/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9450763/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108407244,"identity":"c6c35d36-c69c-48fb-8c83-14cd3d310d8c","added_by":"auto","created_at":"2026-05-04 09:48:55","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":56059,"visible":true,"origin":"","legend":"\u003cp\u003eLASSO Coefficient Profiles\u003c/p\u003e","description":"","filename":"Picture1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9450763/v1/9bce0cfeb79d275aa5de16ee.jpg"},{"id":108493146,"identity":"6ccaae77-6d4e-47c2-b3b7-513df22f9452","added_by":"auto","created_at":"2026-05-05 09:59:29","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":32565,"visible":true,"origin":"","legend":"\u003cp\u003eCross-Validation Curve for LASSO Regression\u003c/p\u003e","description":"","filename":"Picture2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9450763/v1/5984ab1c32c997ebd8e3d012.jpg"},{"id":108407246,"identity":"33d53db4-fc1c-4d8a-b264-f1f19aa07b4a","added_by":"auto","created_at":"2026-05-04 09:48:55","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":207267,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance Comparison of Different Machine Learning Models\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-9450763/v1/b9b064b7b0b6e47b16c080a6.png"},{"id":108407247,"identity":"9b9eddaa-cddf-4ab9-af7f-63cb7cd4f7e7","added_by":"auto","created_at":"2026-05-04 09:48:55","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":467672,"visible":true,"origin":"","legend":"\u003cp\u003eFeature importance and interpretability analysis of the BPNN model for predicting dysphagia in NPC survivors after radiotherapy\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-9450763/v1/dbbe8b538a6261f945eeeb7b.png"},{"id":108804199,"identity":"912f50d9-8968-44fc-9e8c-0228df78f6f2","added_by":"auto","created_at":"2026-05-08 15:17:43","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1177770,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9450763/v1/2e3b44be-dd2b-443c-97a7-903568b1556f.pdf"},{"id":108492595,"identity":"b0edf6c3-7d15-4a97-a46f-b12c8096d0f5","added_by":"auto","created_at":"2026-05-05 09:58:07","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":291959,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterials.docx","url":"https://assets-eu.researchsquare.com/files/rs-9450763/v1/21e4e34312df0e120b91c89b.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"A Machine Learning–Based Prediction Model for Dysphagia After Radiotherapy in Nasopharyngeal Carcinoma Survivors","fulltext":[{"header":"Introduction","content":"\u003cp\u003eNasopharyngeal carcinoma (NPC) is highly prevalent in southern China, with an incidence ranging from 25 to 30 per 100,000 population\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. The adoption of comprehensive treatment strategies, primarily based on intensity-modulated radiation therapy (IMRT) combined with chemotherapy, has increased the 5-year survival rate of NPC patients to over 85%\u003csup\u003e2\u003c/sup\u003e. However, while high-dose radiotherapy effectively eradicates tumors, it inevitably causes damage to surrounding normal tissues, resulting in a range of late complications\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. Among these, dysphagia is one of the most common and severe sequelae, with a reported incidence ranging from 13% to 93.5%\u003csup\u003e4\u003c/sup\u003e. Dysphagia not only leads to physical complications such as aspiration, malnutrition, and aspiration pneumonia, but also significantly impairs patients' psychological well-being and quality of life\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eNotably, dysphagia is characterized by a delayed onset and progressive deterioration, with some survivors experiencing functional decline even 2\u0026ndash;10 years after the completion of radiotherapy\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. However, most existing studies have focused on swallowing function during or shortly after radiotherapy, with limited attention given to the dynamic changes and risk prediction of dysphagia among long-term survivors\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. In clinical practice, early symptoms are often overlooked by patients, and issues such as underdiagnosis and suboptimal management remain prevalent\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. Therefore, systematic assessment and risk prediction of swallowing function in NPC survivors after radiotherapy are of substantial clinical importance for delaying disease progression, reducing hospitalization, and lowering healthcare costs.\u003c/p\u003e \u003cp\u003ePrevious studies have primarily employed traditional statistical methods, such as LR and Cox regression, to identify risk factors for dysphagia\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. However, clinical data are often characterized by nonlinearity, complex interactions among variables, and missing values, which limit the predictive accuracy of conventional approaches\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e. In recent years, machine learning, with its strong capabilities in data mining and self-learning, has demonstrated significant advantages in handling high-dimensional and complex medical data and in developing accurate predictive models\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. For instance, Du et al.\u003csup\u003e11\u003c/sup\u003e applied machine learning techniques to predict the risk of breast cancer\u0026ndash;related lymphedema in 670 patients, achieving promising results. Nevertheless, the application of machine learning in predicting post-radiotherapy dysphagia in NPC remains at an early stage and warrants further investigation and validation.\u003c/p\u003e \u003cp\u003eIn this study, we analyzed factors associated with dysphagia in NPC survivors and employed univariate analysis and least absolute shrinkage and selection operator (LASSO) regression for variable selection. Three predictive models\u0026mdash;logistic regression (LR), backpropagation neural network (BPNN), and classification and regression tree (CART)\u0026mdash;were constructed and compared to identify the optimal model. The aim was to facilitate early identification of high-risk NPC survivors following radiotherapy and to support the implementation of targeted interventions to reduce the incidence of dysphagia.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eResearch design\u003c/h2\u003e \u003cp\u003eThis study employed a cross-sectional design.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eSample and setting in research\u003c/h3\u003e\n\u003cp\u003eA convenience sampling method was used to recruit NPC survivors who attended a tertiary Grade A hospital in Guangdong Province between October 2023 and November 2025. The inclusion criteria were as follows: (1) histopathological confirmation of NPC based on imaging and endoscopic biopsy according to the 8th edition of the AJCC staging system\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e; (2) age\u0026thinsp;\u0026ge;\u0026thinsp;18 years; (3) completion of radiotherapy; (4) provision of informed consent and voluntary participation; and (5) adequate cognitive and communication abilities. The exclusion criteria included: (1) severe cognitive impairment, psychiatric disorders, or other conditions affecting questionnaire completion and communication; and (2) critical illness or the presence of other severe organic diseases.\u003c/p\u003e \u003cp\u003eFor the development of the prediction models, a total of 21 candidate predictors were initially identified based on relevant literature and institutional data\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. According to the events per variable principle, a minimum of 210 positive events (21 \u0026times; 10) was required for model construction. Based on preliminary pilot data, the incidence of dysphagia among NPC survivors was approximately 57%. Therefore, the minimum required sample size was estimated to be 369 cases (210\u0026thinsp;\u0026divide;\u0026thinsp;0.57). Considering a potential invalid response rate of 10%\u0026ndash;20%, a total of 410\u0026ndash;462 participants were deemed necessary. Model performance was evaluated using cross-validation, with the dataset randomly divided into a training set (70%) and a testing set (30%).\u003c/p\u003e\n\u003ch3\u003eData Collection\u003c/h3\u003e\n\u003cp\u003eFace-to-face surveys were conducted by two specially trained nurses when NPC survivors returned for follow-up visits. Participants were informed of the study objectives and instructions for questionnaire completion, and written informed consent was obtained prior to administration. Standardized instructions were provided, and questionnaires were self-administered. For participants who had difficulty completing the questionnaire, trained investigators assisted by recording responses based on the participants' verbal reports with their consent. Each interview lasted approximately 15\u0026ndash;30 minutes and was completed during waiting periods. Disease-related information was supplemented by reviewing the electronic medical record system. All questionnaires were collected and checked on site; those with more than 10% missing data or with patterned responses were excluded as invalid. Data were double-checked by two researchers before entry. A total of 542 questionnaires were distributed, of which 499 were valid, yielding an effective response rate of 92.06%.\u003c/p\u003e\n\u003ch3\u003eEthical Considerations\u003c/h3\u003e\n\u003cp\u003eThis study was approved by the Medical Ethics Committee of the authors' institution (approval number: SZR2022-179), and all participants provided informed consent prior to enrollment.\u003c/p\u003e\n\u003ch3\u003eInstruments\u003c/h3\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eSociodemographic and Clinical Characteristics\u003c/h2\u003e \u003cp\u003eBased on a review of the literature and expert consultation, a self-designed questionnaire was developed by the researchers. It included sociodemographic variables (age, sex, place of residence, marital status, educational level, monthly household income per capita, return-to-work status, and adherence to swallowing training) and disease-related variables (treatment stage, AJCC stage, T/N/M classification, radiotherapy modality, receipt of induction chemotherapy, immunotherapy, and targeted therapy).\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eM.D. Anderson Dysphagia Inventory (MDADI)\u003c/h3\u003e\n\u003cp\u003eThe M.D. Anderson Dysphagia Inventory (MDADI), developed by Chen et al.\u003csup\u003e14\u003c/sup\u003e in 2001, consists of 20 items across four domains: global, emotional, functional, and physical. Each item is rated on a 5-point Likert scale ranging from \"strongly agree\" to \"strongly disagree\" (scored 1\u0026ndash;5), with higher scores indicating better swallowing function and quality of life. In this study, a score\u0026thinsp;\u0026gt;\u0026thinsp;69 was used to define the presence of dysphagia. The \u003cem\u003eCronbach's α\u003c/em\u003e coefficient of the scale is 0.82\u003csup\u003e15\u003c/sup\u003e.\u003c/p\u003e\n\u003ch3\u003evoice index (VHI)\u003c/h3\u003e\n\u003cp\u003eThe voice index (VHI), proposed by Jacobson et al.\u003csup\u003e16\u003c/sup\u003e in 1997, assesses voice-related impairment across three domains: functional, physical, and emotional. It includes 30 items, each scored from 0 (\"never\") to 4 (\"always\"), yielding a total score ranging from 0 to 120. Higher scores indicate greater perceived voice impairment. The \u003cem\u003eCronbach's α\u003c/em\u003e coefficient is 0.86\u003csup\u003e17\u003c/sup\u003e.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eNeck Disability Index (NDI)\u003c/h2\u003e \u003cp\u003eThe Neck Disability Index (NDI), developed by Vernon et al.\u003csup\u003e18\u003c/sup\u003e in 1991, consists of 10 items, each scored from 0 to 5, with a total score ranging from 0 to 50. Higher scores indicate more severe functional impairment. The \u003cem\u003eCronbach's α\u003c/em\u003e coefficient of the scale is \u0026gt;\u0026thinsp;0.81\u003csup\u003e19\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eM.D. Anderson Symptom Inventory\u0026ndash;Head and Neck Module (MDASI-H\u0026amp;N)\u003c/h2\u003e \u003cp\u003eThe MDASI-H\u0026amp;N, developed by Rosenthal et al.\u003csup\u003e20\u003c/sup\u003e, is based on the M.D. Anderson Symptom Inventory and consists of two parts. The first part assesses the severity of 22 symptoms experienced in the past 24 hours, using an 11-point Likert scale ranging from 0 (\"not present\") to 10 (\"as severe as you can imagine\"), with higher scores indicating greater symptom severity. The second part evaluates the extent to which symptoms interfere with daily life and includes 6 items. The Chinese version was translated and validated by Han Yuan et al.\u003csup\u003e21\u003c/sup\u003e in 2010, with \u003cem\u003eCronbach's α\u003c/em\u003e coefficients of 0.877 and 0.835 for the core and module-specific items, respectively.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eGeneralized Anxiety Disorder-7 (GAD-7)\u003c/h2\u003e \u003cp\u003eThe Generalized Anxiety Disorder-7 (GAD-7), developed by Spitzer et al.\u003csup\u003e22\u003c/sup\u003e based on the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV), consists of 7 items rated on a 4-point Likert scale from \"not at all\" to \"nearly every day,\" with total scores ranging from 0 to 21. Cut-off scores of 5, 10, and 15 represent mild, moderate, and severe anxiety, respectively. The \u003cem\u003eCronbach's α\u003c/em\u003e coefficient is 0.949\u003csup\u003e23\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003ePatient Health Questionnaire-9 (PHQ-9)\u003c/h2\u003e \u003cp\u003eThe Patient Health Questionnaire-9 (PHQ-9) uses a 4-point Likert scale (0 = \"not at all\" to 3 = \"nearly every day\"), with total scores ranging from 0 to 27. Cut-off scores of 5, 10, and 15 indicate mild, moderate, and severe depression, respectively\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. In this study, a PHQ-9 score\u0026thinsp;\u0026ge;\u0026thinsp;5 was used as the screening threshold for depression. The \u003cem\u003eCronbach's α\u003c/em\u003e coefficient of the scale is 0.917\u003csup\u003e25\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eTreatment Regimens\u003c/h2\u003e \u003cdiv id=\"Sec16\" class=\"Section3\"\u003e \u003ch2\u003eRadiotherapy\u003c/h2\u003e \u003cp\u003eTarget volume delineation was performed in accordance with the International Commission on Radiation Units and Measurements (ICRU) Reports No. 50. The gross tumor volume (GTV) included the primary tumor and metastatic cervical lymph nodes\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. The clinical target volume (CTV) was individually defined based on the extent of tumor infiltration. Planning target volumes (PTVs) were generated by adding a 3-mm margin to the CTV to account for setup uncertainties.\u003c/p\u003e \u003cp\u003eThe prescribed radiation doses were 70 Gy to the GTV, 66\u0026ndash;70 Gy to involved lymph nodes, 60\u0026ndash;62 Gy to the high-risk CTV, and 54\u0026ndash;56 Gy to the low-risk CTV, delivered in 30\u0026ndash;33 fractions. IMRT and helical tomotherapy (HT) plans were assigned to experienced medical physicists, and all treatment plans were reviewed and approved by senior radiation oncologists. Treatment was delivered using a synchronized integrated boost technique, with one fraction per day, five days per week, over 6\u0026ndash;7 weeks.\u003c/p\u003e \u003cp\u003eIMRT plans were generated using nine coplanar fields. Optimization was performed using the dose-volume optimizer algorithm in the Eclipse treatment planning system (TPS). The plans were delivered using a linear accelerator equipped with a dynamic multileaf collimator system (NOMOS Corporation, Sewickley, PA, USA) in a slice-by-slice arc rotation technique\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eHT plans were generated using a tomotherapy planning system with 6-MV photon beams. The main optimization parameters included a field jaw width of 1.0 cm, a pitch of 0.287, and a modulation factor of 3.8. Dose calculations were performed using a convolution\u0026ndash;superposition algorithm with a fine calculation grid of 0.273 cm \u0026times; 0.273 cm \u0026times; 0.3 cm\u003csup\u003e28\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eChemotherapy\u003c/h2\u003e \u003cp\u003eInduction chemotherapy was administered according to the patient's clinical condition, using platinum-based regimens at 21-day intervals. Common induction chemotherapy (IC) protocols included docetaxel\u0026ndash;cisplatin\u0026ndash;5-fluorouracil (TPF), docetaxel\u0026ndash;cisplatin (TP), cisplatin\u0026ndash;5-fluorouracil (PF), and gemcitabine\u0026ndash;cisplatin (GP). Concurrent chemotherapy during radiotherapy consisted of cisplatin at a dose of 80 or 100 mg/m\u0026sup2; for 2\u0026ndash;3 cycles\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eData analysis\u003c/h2\u003e \u003cp\u003eStatistical analyses were performed using SPSS version 27.0 (IBM Corp., Armonk, NY, USA), and machine learning\u0026ndash;based model development and validation were conducted using Python.\u003c/p\u003e \u003cp\u003eFor descriptive analysis, continuous variables with a non-normal distribution were presented as median (interquartile range, M [P25, P75]) and compared using the Mann\u0026ndash;Whitney U test. Categorical variables were expressed as frequencies (percentages) and compared using the chi-square test or Fisher's exact test when expected cell counts were \u0026lt;\u0026thinsp;5. Normally distributed continuous variables were reported as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation (x̄ \u0026plusmn; s) and compared using independent-samples t-tests.\u003c/p\u003e \u003cp\u003eExploratory factor analysis was performed using maximum variance orthogonal rotation to extract symptom factors. Factors were retained according to the following criteria: eigenvalues\u0026thinsp;\u0026ge;\u0026thinsp;1, inclusion of at least two symptoms per factor, and factor loadings\u0026thinsp;\u0026ge;\u0026thinsp;0.4. When a symptom loaded\u0026thinsp;\u0026ge;\u0026thinsp;0.4 on multiple factors, it was assigned to the factor with the highest loading\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eFor predictive modeling, Python was used to construct three models: LR, BPNN, and CART. Variables identified through univariate LR were further selected using LASSO regression. All features were standardized using Z-score normalization. The optimal regularization parameter (α) was determined via 10-fold cross-validation, and variables with non-zero coefficients were retained as final predictors. The selected features were subsequently used to build multivariable LR, BPNN, and CART models. The dataset was randomly split into a training set (n\u0026thinsp;=\u0026thinsp;349) and a validation set (n\u0026thinsp;=\u0026thinsp;150) at a ratio of 7:3\u003csup\u003e31\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003e(1) The CART model was constructed using the Gini index as the splitting criterion. To prevent overfitting, the maximum tree depth was restricted to 4, and class-weight balancing was applied. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, calibration curves, and Brier scores in both the training and validation sets.\u003c/p\u003e \u003cp\u003e(2) The BPNN model was developed using standardized continuous variables. The network consisted of two hidden layers (30 and 15 neurons, respectively), with ReLU activation functions and the Adam optimizer. L2 regularization and early stopping were applied to reduce overfitting. Feature importance was assessed using permutation importance. Model performance was evaluated using ROC curves (AUC), calibration curves, and Brier scores.\u003c/p\u003e \u003cp\u003e(3) The multivariable LR model was used to identify independent risk factors for dysphagia, with odds ratios (ORs) and 95% confidence intervals (CIs) reported. Model discrimination was assessed using ROC curves (AUC, sensitivity, and specificity), while calibration was evaluated using calibration curves and Brier scores.\u003c/p\u003e \u003cp\u003eFinally, the optimal predictive model for post-radiotherapy dysphagia in NPC survivors was selected based on comprehensive performance evaluation in both the training and validation sets.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003eParticipant Characteristics\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA total of 499 participants were included in this study, among whom 196 survivors developed dysphagia, yielding an overall incidence of 39.3%. In the training set, 143 of 349 survivors experienced dysphagia, corresponding to an incidence of 41.0%. In the validation set, 53 of 150 patients developed dysphagia, with an incidence of 35.3%. There was no statistically significant difference in the incidence of dysphagia between the two groups (\u003cem\u003e\u0026chi;\u0026sup2;\u003c/em\u003e = 1.400, P \u0026gt; 0.05). The mean age of the participants was 45.55 \u0026plusmn; 11.05 years. The proportion of males (63.5%) to females (36.5%) was approximately 2:1. Patients with stage III/IV disease accounted for 80.1% of the sample. The majority of survivors were married, comprising 85.8% of the study population. Detailed sociodemographic and clinical characteristics are presented in Table 1.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1\u003c/strong\u003e General Characteristics of NPC Survivors After Radiotherapy (\u003cem\u003eN\u003c/em\u003e=499)\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eVariables\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003eTraining Set (n=349)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003eValidation Set (n=150)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u003cem\u003e\u0026chi;\u0026sup2;/Z\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u003cem\u003eP\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eVariables\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eSex\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e0.445\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.505\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eMale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e225 (64.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e92 (61.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eFemale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e124 (35.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e58 (38.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eAge, years\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e45 (17)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e43 (13)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e-1.398\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.162\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eReturn to work\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e0.011\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.917\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e153 (43.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e65 (43.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eNo\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e196 (56.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e85 (56.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eResidence\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e0.002\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.965\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eUrban\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e145 (41.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e62 (41.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eRural\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e204 (58.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e88 (58.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eMarital status\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e0.009\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.924\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eMarried\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e50 (14.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e21 (14.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eUnmarried/divorced/widowed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e299 (85.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e129 (86.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eEducation level\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e1.619\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.445\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eMiddle school or below\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e136 (39.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e51 (34.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eHigh school/vocational\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e73 (20.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e38 (25.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eCollege or above\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e140 (40.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e61 (40.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eMonthly income (CNY)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e6.052\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.109\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003e\u0026lt;3000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e144 (41.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e58 (38.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003e3001\u0026ndash;5000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e87 (24.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e36 (24.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003e5001\u0026ndash;10000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e72 (20.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e44 (29.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003e\u0026gt;10000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e46 (13.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e12 (8.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eAdherence to swallowing exercises\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e3.241\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.072\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eNone/partial\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e315 (90.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e127 (84.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eFull\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e34 (9.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e23 (15.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eTreatment stage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e8.100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.044\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003e\u0026lt;1 year\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e191 (54.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e66 (44.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003e1\u0026ndash;3 years\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e74 (21.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e46 (30.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003e3\u0026ndash;5 years\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e44 (12.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e15 (10.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003e\u0026gt;5 years\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e40 (11.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e23 (15.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eAJCC stage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e1.887\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.596\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eI\u0026ndash;II\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e50 (14.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e15 (10.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eIII\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e167 (47.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e78 (52.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eIV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e108 (30.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e47 (31.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eUnknown\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e24 (6.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e10 (6.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eT stage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e0.794\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.851\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eT1\u0026ndash;T2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e43 (12.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e21 (14.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eT3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e225 (64.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e91 (60.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eT4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e73 (20.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e35 (23.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eUnknown\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e8 (2.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e3 (2.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eN stage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e3.718\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.446\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eN0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e28 (8.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e12 (8.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eN1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e119 (34.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e40 (26.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eN2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e104 (29.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e56 (37.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eN3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e90 (25.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e39 (26.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eUnknown\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e8 (2.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e3 (2.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eM stage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e0.468\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.791\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eM0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e327 (93.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e139 (92.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eM1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e14 (4.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e8 (5.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eUnknown\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e8 (2.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e3 (2.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eRadiotherapy technique\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e0.393\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.531\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eIMRT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e320 (91.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e140 (93.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eTOMO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e29 (8.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e10 (6.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eInduction chemotherapy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e1.335\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.248\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e294 (84.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e120 (80.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eNo\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e55 (15.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e30 (20.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eImmunotherapy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e0.774\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.379\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e145 (41.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e56 (37.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eNo\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e204 (58.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e94 (62.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eTargeted therapy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e2.284\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.131\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e97 (27.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e32 (21.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eNo\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e252 (72.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e118 (78.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eDysphagia\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e1.400\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.237\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e143 (41.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e53 (35.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eNo\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e206 (59.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e97 (64.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eVoice index\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e6.00 (12)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e5.50 (12)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e-0.236\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.814\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eDepression\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e4.00 (8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e4.00 (7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e-0.048\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.961\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eAnxiety\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e2.00 (7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e2.00 (7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e-0.140\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.889\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eGastrointestinal/oral symptoms\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e2.45 (3.09)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e2.09 (3.02)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e-0.483\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.629\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eEmotional/fatigue symptoms\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e2.14 (2.93)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e2.14 (2.90)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e-0.235\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.814\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eFunction/quality of life symptoms\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e1.33 (3.09)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e1.33 (3.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e-0.673\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.501\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 189px;\"\u003e\n \u003cp\u003eCervical dysfunction level\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e6.00 (7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 100px;\"\u003e\n \u003cp\u003e6.00 (7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e-0.041\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e0.967\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eUnivariate Analysis of Dysphagia\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePrior to univariate regression analysis, symptom clusters were categorized into three groups based on exploratory factor analysis loadings: the gastrointestinal/oral symptom cluster, emotional/fatigue symptom cluster, and function/quality of life cluster (see Supplementary Table S1 for details). The results of the univariate analysis indicated that multiple variables were significantly associated with the occurrence of dysphagia in the training set (\u003cem\u003ep\u0026nbsp;\u003c/em\u003e\u0026lt; 0.05). Among these, sociodemographic factors included return-to-work status, residence, educational level, monthly income, and age. Clinical factors comprised AJCC stage, N stage, and receipt of immunotherapy and targeted therapy. Psychological factors included levels of depression and anxiety. Symptom- and function-related variables included voice index, gastrointestinal/oral symptom cluster, emotional/fatigue symptom cluster, function/quality of life cluster, and cervical dysfunction level. Detailed results are presented in Supplementary Table S2.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eLASSO Regression Analysis of Dysphagia\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eLASSO regression was performed to further select predictive variables for post-radiotherapy dysphagia in NPC survivors. Variables identified as statistically significant in the univariate analysis were included in the LASSO model. As the regularization parameter (\u003cem\u003e\u0026lambda;\u003c/em\u003e) increased, the mean squared error (MSE) initially decreased and then increased. At the optimal value of \u003cem\u003e\u0026lambda;\u003c/em\u003e, the regression coefficients of residence and anxiety were shrunk to zero and thus excluded from the final model. Ultimately, 14 variables with relatively stable non-zero coefficients were retained, suggesting that they may be important predictors of dysphagia and were subsequently included in the risk prediction models. Detailed results are presented in Figures 1 and 2.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFigure 1\u003c/strong\u003e LASSO Coefficient Profiles\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFigure 2\u003c/strong\u003e Cross-Validation Curve for LASSO Regression\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eModel Development and Evaluation\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn this study, three predictive models\u0026mdash;CART, BPNN, and LR\u0026mdash;were developed to estimate the risk of post-radiotherapy dysphagia in patients with NPC. Model performance was evaluated in both the training and validation sets, with discrimination assessed using ROC curves and the AUC. In the training set, the AUC values for the CART, BPNN, and LR models were 0.824, 0.783, and 0.824, respectively. In the validation set, the corresponding AUC values were 0.666, 0.776, and 0.753, indicating overall good predictive performance (see Figure 3 and Supplementary Figures S1\u0026ndash;S3).\u003c/p\u003e\n\u003cp\u003eComparative analysis showed that CART and LR achieved the highest discrimination in the training set (AUC = 0.824). However, in the validation set, the BPNN model demonstrated superior robustness, with the highest AUC (0.776) and accuracy (0.720) among the three models. In terms of specificity, both the BPNN and LR models exhibited high performance in the validation set (0.845 and 0.825, respectively), indicating strong capability in correctly identifying patients at risk of dysphagia. Overall, the BPNN model demonstrated the best comprehensive performance and may serve as a promising tool for clinical application. Detailed results are presented in Table 3.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4\u0026nbsp;\u003c/strong\u003eComparison of predictive performance among the three models\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"3\" cellpadding=\"0\" width=\"99%\" class=\"fr-table-selection-hover\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 18px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTraining Set\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eValidation Set\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 18px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMetrics\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCART\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBPNN\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLR\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCART\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBPNN\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLR\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 18px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUC\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19px;\"\u003e\n \u003cp\u003e0.824\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e0.783\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e0.824\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22px;\"\u003e\n \u003cp\u003e0.666\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e0.776\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e0.753\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 18px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19px;\"\u003e\n \u003cp\u003e0.777\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e0.745\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e0.777\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22px;\"\u003e\n \u003cp\u003e0.653\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e0.720\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e0.700\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 18px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSensitivity\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19px;\"\u003e\n \u003cp\u003e0.748\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e0.587\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e0.664\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22px;\"\u003e\n \u003cp\u003e0.623\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e0.528\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e0.472\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 18px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecificity\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19px;\"\u003e\n \u003cp\u003e0.796\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e0.854\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e0.854\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22px;\"\u003e\n \u003cp\u003e0.670\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e0.845\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e0.825\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 18px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBrier Score\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e0.181\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e0.159\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e0.180\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e0.186\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eNotes: AUC, area under the curve; CI, confidence interval. DeLong test used for comparisons.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eImportance Ranking of Predictors in the Neural Network Model\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe results (Figure 4) indicated that, as measured by the decrease in AUC, depression was the most important predictor (0.0650), substantially exceeding all other variables. This was followed by the voice index (0.0403) and the gastrointestinal/oral symptom cluster (0.0209). In addition, N stage (N2: 0.0128; N3: 0.0105), Cervical dysfunction level (0.0115), and targeted therapy (0.0077) also demonstrated notable predictive value. In contrast, sociodemographic variables such as age and income showed relatively low importance, with some even exhibiting negative contributions, suggesting limited impact on improving model performance. The SHAP analysis results in the testing set were generally consistent with these findings, highlighting the dominant role of symptom burden and psychological factors in the model. Furthermore, the SHAP summary plot illustrated the directional effects of predictors on model output: higher levels of depression, greater symptom burden, and more severe cervical dysfunction were all associated with an increased risk of dysphagia. By comparison, certain sociodemographic variables (e.g., income level and return-to-work status) showed weaker and more variable effects across individuals.\u003c/p\u003e\n\u003cp\u003eOverall, the predictive performance of the model was primarily driven by psychological status, symptom burden, and functional impairment, whereas traditional sociodemographic factors contributed relatively little. These findings suggest that post-radiotherapy dysphagia in NPC survivors is characterized by complex, multidimensional interactions. Detailed results are presented in Supplementary Figure S4.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003e \u003cb\u003eThe high incidence of dysphagia among NPC survivors highlights the need to establish a comprehensive, longitudinal monitoring system.\u003c/b\u003e \u003c/p\u003e \u003cp\u003eIn this study, the incidence of post-radiotherapy dysphagia among NPC survivors was 39.3%, which is generally consistent with the 42.8% reported by Li et al.\u003csup\u003e7\u003c/sup\u003e. Currently, IMRT has replaced conventional radiotherapy as the standard treatment modality for NPC. Through precise spatial dose modulation, IMRT effectively reduces radiation exposure to surrounding normal tissues, thereby lowering the incidence of dysphagia compared with conventional techniques\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. However, due to the unique anatomical location of the nasopharynx\u0026mdash;adjacent to swallowing-related muscles and cranial nerves\u0026mdash;radiation-induced damage remains difficult to completely avoid.\u003c/p\u003e \u003cp\u003eFrom a pathophysiological perspective, radiation-induced injury to cranial nerves may result in paralysis and sensory impairment of the pharyngeal and laryngeal muscles, leading to delayed swallowing reflexes and aspiration. In addition, direct damage to the swallowing musculature can cause fibrosis and atrophy, ultimately impairing swallowing efficiency\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. Notably, dysphagia is characterized by a delayed onset and progressive worsening, with some patients developing symptoms months or even years after the completion of radiotherapy\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. McDowell et al.\u003csup\u003e32\u003c/sup\u003e reported in a long-term follow-up study of NPC survivors treated with IMRT that a considerable proportion of patients continued to experience swallowing and chewing difficulties even 7.5 years post-treatment, which were closely associated with emotional distress, including depression and anxiety. Therefore, in clinical practice, it is essential to establish a dynamic, longitudinal assessment framework beginning at the initiation of radiotherapy. Regular monitoring of swallowing function and structured follow-up should be implemented to facilitate early identification of high-risk individuals and enable timely, individualized interventions, ultimately reducing the incidence of dysphagia and improving long-term quality of life.\u003c/p\u003e \u003cdiv id=\"Sec26\" class=\"Section2\"\u003e \u003ch2\u003eComparison of the predictive performance and clinical applicability of different machine learning models\u003c/h2\u003e \u003cp\u003eThis study, based on clinical data from 499 NPC patients after radiotherapy, systematically compared the predictive performance of three machine learning models (CART, BPNN, and LR) for post-radiotherapy dysphagia. The results demonstrated that the BPNN model exhibited the best robustness in the validation set, with a high specificity of 0.845, indicating its strong ability to correctly identify patients without dysphagia. In contrast, the LR and CART models showed relatively higher discrimination in the training set, suggesting potential limitations in generalizability. Overall, the BPNN model demonstrated superior clinical applicability and potential value for identifying high-risk populations. Srinivasan et al.\u003csup\u003e33\u003c/sup\u003e, in a recent review, reported that machine learning models, by integrating multimodal data, exhibit superior discriminative ability compared with traditional models in predicting treatment-related toxicities such as dysphagia and voice impairment in head and neck cancer patients. The BPNN model can automatically capture nonlinear relationships and complex interactions among variables, and continuously optimize model weights through the backpropagation algorithm, thereby improving predictive accuracy. This finding is consistent with the results reported by Nicol et al.\u003csup\u003e34\u003c/sup\u003e. A large-scale study involving 1,685 head and neck cancer patients\u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e also demonstrated that mixed-effect models incorporating dosimetric factors achieved stable discriminative performance in predicting dysphagia (AUC\u0026thinsp;=\u0026thinsp;0.77\u0026ndash;0.84).\u003c/p\u003e \u003cp\u003eHowever, the CART model showed evidence of overfitting in the validation set (AUC: 0.824 in the training set vs. 0.666 in the validation set), indicating limited generalizability. In contrast, the BPNN model demonstrated more stable predictive performance and may be better suited for clinical decision-support systems, although its reliability should still be validated in specific clinical contexts.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec27\" class=\"Section2\"\u003e \u003ch2\u003eMultidimensional determinants of dysphagia risk in NPC survivors\u003c/h2\u003e \u003cp\u003eThis study identified depression as the most important predictor in the BPNN-based feature importance analysis, highlighting the critical role of psychological factors in swallowing rehabilitation. Wang et al.\u003csup\u003e36\u003c/sup\u003e reported that 44.2% of 773 post-radiotherapy NPC survivors experienced mild to moderate depressive symptoms. Similarly, McDowell et al.\u003csup\u003e32\u003c/sup\u003e found that among NPC survivors treated with IMRT, the prevalence of depression and anxiety was 25% and 37%, respectively, both of which were significantly negatively associated with quality of life. Dysphagia-related eating difficulties, nutritional risk, and social withdrawal may directly lead to or exacerbate depressive symptoms. In turn, impaired swallowing function may cause patients to avoid public eating, alter dietary patterns, and experience disruption in daily life routines, with some individuals adopting maladaptive coping strategies that further reduce quality of life\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. Moreover, depression may amplify the subjective perception of dysphagia, pain, and fatigue, and may also influence physiological recovery through neuroendocrine and inflammatory pathways, thereby creating a bidirectional \"symptom\u0026ndash;psychology\" vicious cycle that further worsens swallowing dysfunction\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. Therefore, rehabilitation strategies focusing solely on swallowing function may be insufficient, and integrated approaches combining swallowing rehabilitation with psychological intervention are warranted in clinical practice.\u003c/p\u003e \u003cp\u003eThe voice index ranked as the second most important predictor, suggesting a close anatomical and functional association between voice and swallowing dysfunction, which is consistent with the findings of Wentzel et al.\u003csup\u003e37\u003c/sup\u003e. Swallowing and phonatory functions share common anatomical structures, including the tongue, soft palate, pharynx, and larynx, and damage to these structures or their neural control may simultaneously impair both functions\u003csup\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e. Portas et al.\u003csup\u003e39\u003c/sup\u003e demonstrated in patients after laryngeal cancer surgery that voice index can indicate the presence of swallowing impairment. Therefore, systematic collection of patient-reported voice outcomes (e.g., VHI) during clinical follow-up may not only facilitate voice rehabilitation management but also provide important clues for early identification of dysphagia risk.\u003c/p\u003e \u003cp\u003eThis study further found that the gastrointestinal/oral symptom cluster was highly influential in the model, indicating that dysphagia is not an isolated event but is closely associated with multiple co-occurring symptoms. Previous studies have confirmed that symptoms in NPC survivors tend to cluster, with xerostomia, pain, taste alteration, and dysphagia interacting synergistically to impair eating behavior and quality of life\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e. A cross-sectional study by Tam et al.\u003csup\u003e41\u003c/sup\u003e involving 211 survivors further supported this view, showing that nearly all patients experienced late radiation-related toxicities, and more than three-quarters reported at least four moderate-to-severe symptoms. Dysphagia and xerostomia were among the top five most common symptoms, and symptom burden was significantly positively correlated with unresolved symptom distress, with no improvement over time. From a pathophysiological perspective, radiation-induced damage to salivary glands and pharyngeal muscles leads to impaired bolus formation and pharyngeal residue, thereby increasing the risk of dysphagia\u003csup\u003e\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e\u003c/sup\u003e. These findings suggest that post-radiotherapy dysphagia management in NPC survivors should move beyond a single-symptom approach toward an integrated strategy combining symptom cluster monitoring and functional rehabilitation.\u003c/p\u003e \u003cp\u003eIn addition, cervical dysfunction and N stage remained important predictors, suggesting that radiation-induced structural damage and tumor burden play fundamental roles in dysphagia development. Post-radiotherapy fibrosis and atrophy of cervical muscles lead to restricted neck mobility, weakened laryngeal musculature, impaired pharyngoesophageal movement, and reduced neural reflex function, ultimately resulting in poor coordination of swallowing and mastication\u003csup\u003e\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e. Although IMRT has improved survival outcomes, more intensive treatment regimens are also associated with a higher incidence of dysphagia\u003csup\u003e\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u003c/sup\u003e. Moreover, higher N stage generally indicates a wider irradiation field and more extensive tissue involvement. Reduced regenerative capacity, reactive exudation, tissue sclerosis, and fibrosis within the irradiated field may lead to trismus and restricted soft tissue mobility, further increasing dysphagia risk\u003csup\u003e\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e,\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e. Targeted therapy, while enhancing anti-tumor effects through inhibition of angiogenesis and immune-mediated tumor suppression\u003csup\u003e\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e\u003c/sup\u003e, may also increase the incidence of skin reactions and oral mucositis, thereby exacerbating injury to the orofacial and cervical musculature.\u003c/p\u003e \u003cp\u003eFinally, SHAP analysis based on the validation set further elucidated both the magnitude and directionality of variable contributions. Depression, symptom cluster burden, and cervical dysfunction consistently exerted positive effects on dysphagia risk, indicating that higher levels of these factors were associated with increased risk across individuals. In contrast, certain socioeconomic variables (e.g., income level and return-to-work status) showed heterogeneous effects across individuals, suggesting context-dependent influences. Compared with traditional statistical methods, SHAP enables quantification of marginal contributions at the individual level and reveals potential nonlinear relationships and interaction effects, thereby enhancing model interpretability and clinical applicability \u003csup\u003e\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003eLimitations\u003c/h2\u003e \u003cp\u003eDespite exploring the determinants of dysphagia from a multidimensional perspective, several limitations should be acknowledged. First, future studies should conduct external validation using multicenter, large-scale datasets to enhance the generalizability and clinical applicability of the model. Second, incorporating imaging features, biological markers, and objective functional assessments (e.g., videofluoroscopic swallowing study or electromyography) may help develop more refined predictive models and further improve predictive accuracy. Third, complex interactions and potential causal relationships among variables may exist. Future studies employing causal inference methods or longitudinal designs are needed to elucidate the dynamic evolution of psychological factors, symptom clusters, and functional impairment in the development of dysphagia. Finally, based on the key risk factors identified in this study, interventional research should be conducted to evaluate the effectiveness of integrated psychological support, symptom management, and functional rehabilitation strategies in improving swallowing outcomes, thereby enhancing the long-term quality of life of NPC survivors.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eBased on clinical data from 499 patients with NPC after radiotherapy, this study developed dysphagia risk prediction models using three machine learning algorithms, including LR, BPNN and CART. The results showed that the incidence of post-radiotherapy dysphagia was 39.3%. The BPNN model demonstrated the best predictive robustness and relatively high specificity in the validation set, suggesting its potential value for clinical risk stratification and screening. SHAP analysis further indicated that depression, emotional/fatigue symptom clusters, and cervical dysfunction were the main predictive factors. Based on these findings, the management of post-radiotherapy dysphagia in NPC survivors should move beyond a sole focus on structural injury and shift toward a multidimensional risk stratification strategy that integrates psychological status, symptom clusters, and cervical functional training, in order to achieve early identification, precise intervention, and improved long-term quality of life.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eConflict of interest\u003c/h2\u003e \u003cp\u003eThe authors made no disclosures.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eSun Yat-sen University Cancer Center Yue-Qin Nursing Research and Innovation Fund, YQ2025001-A, YQ2025007-B.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eYan He, Hui Zhao, Yongjiao Kang contributed equally to this work and should be considered co-first authors. Yan He, Hui Zhao, Yongjiao Kang were responsible for the study conception, design, and manuscript drafting. Jiajie Xu, Yanling Wen contributed to data collection and management. Jia Li, Yuying Fan and Wen Hu provided critical revisions and guidance on study methodology, statistical analysis, and interpretation of results. All authors read and approved the final manuscript. Jia Li, Yuying Fan, and Wen Hu are joint corresponding authors.\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e \u003cp\u003eNone.\u003c/p\u003e\u003ch2\u003eData availability statements\u003c/h2\u003e\u003cp\u003eThe datasets used and/or analysed during the current study available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eChen Y-P, Chan ATC, Le Q-T, Blanchard P, Sun Y, Ma J. Nasopharyngeal carcinoma. Lancet. 2019;394:64\u0026ndash;80.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWei X, Chen B, Wang Z, Zhao P, Duan X. Nasopharyngeal cancer risk assessment by country or region worldwide from 1990 to 2019. BMC Public Health. 2024;24:1931.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYao J-J, Zhang F, Gao T-S, Zhang W-J, Lawrence WR, Zhu B-T, et al. Survival impact of radiotherapy interruption in nasopharyngeal carcinoma in the intensity-modulated radiotherapy era: A big-data intelligence platform-based analysis. Radiother Oncol. 2019;132:178\u0026ndash;87.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFong R, Ward EC, Rumbach AF. Dysphagia after chemo-radiation for nasopharyngeal cancer: A scoping review. World J Otorhinolaryngol Head Neck Surg. 2020;6:10\u0026ndash;24.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcCarty EB, Chao TN. Dysphagia and Swallowing Disorders. Med Clin North Am. 2021;105:939\u0026ndash;54.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSuiter DM, Gosa MM. \u003cem\u003eAssessing and Treating Dysphagia: A Lifespan Perspective\u003c/em\u003e. Thieme: Stuttgart; 2019.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi H, Li L, Huang X, Li Y, Zou T, Zhuo X, et al. Radiotherapy-induced dysphagia and its impact on quality of life in patients with nasopharyngeal carcinoma. Strahlenther Onkol. 2019;195:457\u0026ndash;67.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLangendijk JA, Doornaert P, Rietveld DHF, Verdonck-de Leeuw IM, Leemans CR, Slotman BJ. A predictive model for swallowing dysfunction after curative radiotherapy in head and neck cancer. Radiother Oncol. 2009;90:189\u0026ndash;95.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAtias D, Ashri S, Goldbourt U, Benyamini Y, Gilad-Bachrach R, Hasin T, et al. Machine learning in epidemiology: An introduction, comparison with traditional methods, and a case study of predicting extreme longevity. Ann Epidemiol. 2025;110:23\u0026ndash;33.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ede Vette SPM, Neh H, van der Hoek L, MacRae DC, Chu H, Gawryszuk A, et al. Deep learning NTCP model for late dysphagia after radiotherapy for head and neck cancer patients based on 3D dose, CT and segmentations. Radiother Oncol. 2025;213:111169.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDu J, Yang J, Yang Q, Zhang X, Yuan L, Fu B. Comparison of machine learning models to predict the risk of breast cancer-related lymphedema among breast cancer survivors: a cross-sectional study in China. Front Oncol. 2024;14:1334082.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePan JJ, Ng WT, Zong JF, Chan LLK, O\u0026rsquo;Sullivan B, Lin SJ, et al. Proposal for the 8th edition of the AJCC/UICC staging system for nasopharyngeal cancer in the era of intensity-modulated radiotherapy. Cancer. 2016;122:546\u0026ndash;58.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRiley RD, Ensor J, Snell KIE, Harrell FE, Martin GP, Reitsma JB, et al. Calculating the sample size required for developing a clinical prediction model. BMJ. 2020;368:m441.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen AY, Frankowski R, Bishop-Leone J, Hebert T, Leyk S, Lewin J, et al. The development and validation of a dysphagia-specific quality-of-life questionnaire for patients with head and neck cancer: the M. D. Anderson dysphagia inventory. Arch Otolaryngol Head Neck Surg. 2001;127:870\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYee K, Wong SM, Teo I, Loy J, Roche E, Tan YP, et al. Validity and reliability of the MD Anderson dysphagia inventory in English and Chinese in head and neck cancer patients. Asia Pac J Clin Oncol. 2020;16:372\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJohnson A, Jacobson B, Grywalski C, Silbergleit A, Jacobson G, Benninger M. The Voice Handicap Index (VHI): Development and Validation. Am J Speech-Language Pathol. 1997;6:66\u0026ndash;70.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXu W, Li H-Y, Hu R, Hu H-Y, Hou L-Z, Zhang L, et al. Analysis of reliability and validity of the Chinese version of voice handicap index (VHI). Chin J Otorhinolaryngol Head Neck Surg. 2008;43:670\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVernon H, Mior S. The Neck Disability Index: a study of reliability and validity. J Manipulative Physiol Ther. 1991;14:409\u0026ndash;15.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSaltychev M, Pylk\u0026auml;s K, Karklins A, Juhola J. Psychometric properties of neck disability index - a systematic review and meta-analysis. Disabil Rehabil. 2024;46:5415\u0026ndash;31.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRosenthal DI, Mendoza TR, Chambers MS, Burkett VS, Garden AS, Hessell AC, et al. The M. D. Anderson symptom inventory-head and neck module, a patient-reported outcome instrument, accurately predicts the severity of radiation-induced mucositis. Int J Radiat Oncol Biol Phys. 2008;72:1355\u0026ndash;61.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHan Y, Zhang MF, Zhang JE. Symptom distress during and after radiation therapy in patients with nasopharyngeal carcinoma[J]. Chin J Nurs. 2010;45(07):626\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSpitzer RL, Kroenke K, Williams JBW, L\u0026ouml;we B. A brief measure for assessing generalized anxiety disorder: the GAD-7. Arch Intern Med. 2006;166:1092\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGong Y, Zhou H, Zhang Y, Zhu X, Wang X, Shen B, et al. Validation of the 7-item Generalized Anxiety Disorder scale (GAD-7) as a screening tool for anxiety among pregnant Chinese women. J Affect Disord. 2021;282:98\u0026ndash;103.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSpitzer RL, Kroenke K, Williams JB. Validation and utility of a self-report version of PRIME-MD: the PHQ primary care study. Primary Care Evaluation of Mental Disorders. Patient Health Questionnaire JAMA. 1999;282:1737\u0026ndash;44.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang W, Bian Q, Zhao Y, Li X, Wang W, Du J, et al. Reliability and validity of the Chinese version of the Patient Health Questionnaire (PHQ-9) in the general population. Gen Hosp Psychiatry. 2014;36:539\u0026ndash;44.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Q-F, Zhou K-B, Chen Z-J, Long Y-B. Pathogenesis and progress of rehabilitation treatment of common complications after radiotherapy for nasopharyngeal carcinoma. Chin J Otorhinolaryngol - Skull Base Surg. 2023;29:102\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Q, Yao W-Y, Xu P-X, Fang Y-W, Lin Z-Y, Li Z-T, et al. Comparison of radiation-induced temporal lobe injury after helical tomotherapy versus conventional intensity-modulated radiotherapy for nasopharyngeal carcinoma: a propensity score-matched cohort study. Radiat Oncol. 2026;21:23.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi W-F, Li Y-Q, Chen L, Zhang Y, Guo R, Zhang F, et al. Propensity-matched analysis of three different chemotherapy sequences in patients with locoregionally advanced nasopharyngeal carcinoma treated using intensity-modulated radiotherapy. BMC Cancer. 2015;15:810.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee N, Harris J, Garden AS, Straube W, Glisson B, Xia P, et al. Intensity-modulated radiation therapy with or without chemotherapy for nasopharyngeal carcinoma: radiation therapy oncology group phase II trial 0225. J Clin Oncol. 2009;27:3684\u0026ndash;90.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBeam AL, Kohane IS. Big Data and Machine Learning in Health Care. JAMA. 2018;319:1317\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHammer MJ, Harris C, Pozzar R, Snowberg K, Paul SM, Cooper BA, et al. Symptom Clusters and Symptom Burden in Cancer Survivors. Cancer Med. 2026;15:e71653.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcDowell LJ, Rock K, Xu W, Chan B, Waldron J, Lu L, et al. Long-Term Late Toxicity, Quality of Life, and Emotional Distress in Patients With Nasopharyngeal Carcinoma Treated With Intensity Modulated Radiation Therapy. Int J Radiat Oncol Biol Phys. 2018;102:340\u0026ndash;52.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSrinivasan Y, Liu A, Rameau A. Machine learning in the evaluation of voice and swallowing in the head and neck cancer patient. Curr Opin Otolaryngol Head Neck Surg. 2024;32:105\u0026ndash;12.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNicol AJ. Multi-omic prediction of severe acute treatment-induced oral mucositis and dysphagia in nasopharyngeal carcinoma patients. 2025\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://theses.lib.polyu.edu.hk/handle/200/13701\u003c/span\u003e\u003cspan address=\"https://theses.lib.polyu.edu.hk/handle/200/13701\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Accessed 14 April 2026.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWordenskjold Stougaard S, Zukauskaite R, R\u0026ouml;ttger R, Laugaard Lorenzen E, Lukas Konrad M, Long Krogh S, et al. Impact of GTV-CTV margin and other predictors on radiation-induced dysphagia in head and neck cancer patients from DAHANCA group. Acta Oncol. 2025;64:1253\u0026ndash;61.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang S, Lai L, Deng Y, Zhang Y, Cai Y, Huang W, et al. The mediating role of resilience in the relationships between posttraumatic growth and depression as well as anxiety among survivors of nasopharyngeal carcinoma: a cross-sectional study. BMC Psychol. 2025;13:768.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWentzel A, Mohamed ASR, Naser MA, van Dijk LV, Hutcheson K, Moreno AM, et al. Multi-organ spatial stratification of 3-D dose distributions improves risk prediction of long-term self-reported severe symptoms in oropharyngeal cancer patients receiving radiotherapy: development of a pre-treatment decision support tool. Front Oncol. 2023;13:1210087.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDos Santos KW, da Cunha Rodrigues E, Rech RS, da Ros Wendland EM, Neves M, Hugo FN, et al. Using Voice Change as an Indicator of Dysphagia: A Systematic Review. Dysphagia. 2022;37:736\u0026ndash;48.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePortas JG, Queija D dos, Arine S, Ferreira LP, Dedivitis AS, Lehn RA. Voice and swallowing disorders: functional results and quality of life following supracricoid laryngectomy with cricohyoidoepiglottopexy. Ear Nose Throat J. 2009;88:E23\u0026ndash;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi J, He Y, Xiao M-F, Huang X-T, Fan Y-Y, Lv X, Qin H-Y. Network analysis of symptom clusters and Quality of Life in Nasopharyngeal Carcinoma survivors. Chin Nurs Manage. 2025;25:1472\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTam VCW, Ching JCF, Yip SST, Kwong VHY, Chan CPL, Wong KCW, et al. Examining patient-reported late toxicity and its association with quality of life and unmet need for symptom management among nasopharyngeal cancer survivors: a cross-sectional survey. Front Oncol. 2024;14:1378973.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eG\u0026oacute;mez \u0026Aacute;, Garc\u0026iacute;a-Chabur MA, Pe\u0026ntilde;aranda D, G\u0026oacute;mez-Mendoza A, Forero JC. Chemotherapy/Radiotherapy-Induced Dysphagia in Head and Neck Tumors: A Challenge for Otolaryngologists in Low- to Middle-Income Countries. Dysphagia. 2025;40:515\u0026ndash;27.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGelblum D, Wolden S, Schupak K, Lee N. Neck Spasms as a Late Effect of Intensity Modulated Radiation Therapy (IMRT) for Head and Neck Cancer. Int J Radiat Oncol Biol Phys. 2011;81:S550.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYou R, Hua Y-J, Liu Y-P, Yang Q, Zhang Y-N, Li J-B, et al. Concurrent Chemoradiotherapy with or without Anti-EGFR-Targeted Treatment for Stage II-IVb Nasopharyngeal Carcinoma: Retrospective Analysis with a Large Cohort and Long Follow-up. Theranostics. 2017;7:2314\u0026ndash;24.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCarmignani I, Locatello LG, Desideri I, Bonomo P, Olmetto E, Livi L, et al. Analysis of dysphagia in advanced-stage head-and-neck cancer patients: impact on quality of life and development of a preventive swallowing treatment. Eur Arch Otorhinolaryngol. 2018;275:2159\u0026ndash;67.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZheng Y, Han F, Xiao W, Xiang Y, Lu L, Deng X, et al. Analysis of late toxicity in nasopharyngeal carcinoma patients treated with intensity modulated radiation therapy. Radiat Oncol. 2015;10:17.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBonner JA, Harari PM, Giralt J, Azarnia N, Shin DM, Cohen RB, et al. Radiotherapy plus cetuximab for squamous-cell carcinoma of the head and neck. N Engl J Med. 2006;354:567\u0026ndash;78.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLinardatos P, Papastefanopoulos V, Kotsiantis S, Explainable AI. A Review of Machine Learning Interpretability Methods. Entropy (Basel). 2020;23:18.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Nasopharyngeal carcinoma, Radiotherapy, Dysphagia, Risk prediction model, Machine learning","lastPublishedDoi":"10.21203/rs.3.rs-9450763/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9450763/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eDysphagia is a common long-term complication in patients with nasopharyngeal carcinoma following radiotherapy, significantly impairing nutritional status and quality of life. Therefore, early identification of high-risk individuals and timely intervention are of critical importance. However, there is currently a lack of effective risk prediction tools that integrate multidimensional symptom profiles with clinical characteristics. This study aims to develop a machine learning\u0026ndash;based predictive model for dysphagia and to evaluate its predictive performance, thereby providing a reference for early identification and prevention strategies.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eA total of 499 nasopharyngeal carcinoma survivors who attended a tertiary cancer hospital in Guangzhou between October 2023 and November 2025 were included in this study. Participants were randomly divided into a training set (n\u0026thinsp;=\u0026thinsp;349) and a testing set (n\u0026thinsp;=\u0026thinsp;150) at a ratio of 7:3. Data on demographic characteristics, clinical variables, and patient-reported outcomes were collected, including the M.D. Anderson Dysphagia Inventory (MDADI), Voice Handicap Index (VHI), M.D. Anderson Symptom Inventory\u0026ndash;Head and Neck Module (MDASI-H\u0026amp;N), Generalized Anxiety Disorder-7 (GAD-7), and Patient Health Questionnaire-9 (PHQ-9). Predictor variables were selected using univariate analysis and least absolute shrinkage and selection operator (LASSO) regression. Three predictive models were subsequently developed: logistic regression (LR), backpropagation neural network (BPNN), and classification and regression tree (CART). Model discrimination was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. Calibration curves, Brier scores, and Shapley Additive Explanations (SHAP) were further applied to assess model performance and interpretability.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe incidence of post-radiotherapy dysphagia among nasopharyngeal carcinoma survivors was 39.3%. A total of 14 predictive variables were ultimately selected. In the training set, both the LR and CART models achieved an AUC of 0.824, outperforming the BPNN model (AUC\u0026thinsp;=\u0026thinsp;0.783). In the testing set, the BPNN model demonstrated the best performance (AUC\u0026thinsp;=\u0026thinsp;0.776, accuracy\u0026thinsp;=\u0026thinsp;0.720), exceeding that of the LR model (AUC\u0026thinsp;=\u0026thinsp;0.666) and the CART model (AUC\u0026thinsp;=\u0026thinsp;0.753). SHAP analysis indicated that depression, voice handicap index, gastrointestinal/oral symptom cluster, tumor stage, N stage, neck dysfunction, targeted therapy, and 3\u0026ndash;5 years post-radiotherapy were the most influential contributors to the BPNN model.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eThe BPNN model demonstrated good stability and generalizability in predicting post-radiotherapy dysphagia in nasopharyngeal carcinoma survivors, and may facilitate early identification of high-risk patients and the implementation of targeted clinical interventions.\u003c/p\u003e","manuscriptTitle":"A Machine Learning–Based Prediction Model for Dysphagia After Radiotherapy in Nasopharyngeal Carcinoma Survivors","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-04 09:48:50","doi":"10.21203/rs.3.rs-9450763/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"175de79d-d4c9-4686-929d-0b0938b1f06c","owner":[],"postedDate":"May 4th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Withdrawn","date":"2026-05-08T06:45:30+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-05-08T06:56:14+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-04 09:48:50","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9450763","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9450763","identity":"rs-9450763","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}