Establishment and Validation of a Predictive Model for Radiation-Associated Aspiration Pneumonia in Patients with Radiation-Induced Dysphagia after Nasopharyngeal Carcinoma

Introduction Radiotherapy for patients with head and neck cancers raises their risk of aspiration pneumonia-related death. We aimed to develop and validate a model to predict radiation-associated aspiration pneumonia (RAP) among patients with dysphagia after radiotherapy for nasopharyngeal carcinoma (NPC). Materials and Methods A total of 453 dysphagic patients with NPC were retrospectively recruited from Sun Yat-Sen Memorial Hospital from January 2012 to January 2018. Patients were randomly divided into training cohort (n = 302) and internal validation cohort (n = 151) at a ratio of 2 : 1. The concordance index (C-index) and calibration curve were used to evaluate the accuracy and discriminative ability of this model. Moreover, decision curve analysis was performed to evaluate the net clinical benefit. The results were externally validated in 203 dysphagic patients from the First People's Hospital of Foshan. Results Derived from multivariable analysis of the training cohort, four independent factors were introduced to predict RAP, including Kubota water drinking test grades, the maximum radiation dose of lymph node gross tumor volume (Dmax of the GTVnd), neutrophil count, and erythrocyte sedimentation rate (ESR). The nomogram showed favorable calibration and discrimination regarding the training cohort, with a C-index of 0.749 (95% confidence interval (CI), 0.681 to 0.817), which was confirmed by the internal validation cohort (C-index 0.743; 95% CI, 0.669 to 0.818) and the external validation cohort (C-index 0.722; 95% CI, 0.606 to 0.838). Conclusions Our study established and validated a simple nomogram for RAP among patients with dysphagia after radiotherapy for NPC.


Introduction
Radiotherapy-induced dysphagia, with an incidence of 5.7-37.3% in nasopharyngeal carcinoma (NPC) patients [1], usually results in severe pneumonia [2,3]. Deteriorating swallowing function indicates an increased risk of radiotherapy-associated aspiration pneumonia (RAP) [4], almost twice as high as that of nondysphagia patients [5]. The burden brought by RAP could be tremendous. Not only does it prolong hospital stays and ventilatory support in ICU but severely affects patients' quality of life and even be life-threating [6]. Chen et al. reported 43.9% (18/41) patients needed ventilatory support in ICU and 17.1% (7/41) died for aspiration pneumonia postradiotherapy [7]. In addition, head and neck cancer patients are at a higher risk of RAP-related death [8]. RAP accounts for 34.6% of noncancer-related deaths in patients with NPC, increasing financial burden on the medical system [9]. However, RAP dose not garner sufficient attention. There are some available treatments for dysphagic patients to prevent RAP, such as nasogastric feeding tube and gastrostomy. The limitation is also obvious, as nasogastric feeding tube is only valid for 1 month, and it is costly and inconvenient to replace gastric tube, while gastrostomy is an invasive treatment. Thus, early screening patients with high risk and follow ups could help preventive approach implement, minimize occurrence of RAP, and provide precautions.
Nomogram serves as a reliable tool to quantify risk for various diseases [10]. However, nomogram for predicting RAP after radiotherapy in dysphagic patients was yet to be developed. In this study, we sought to develop a useful and practical nomogram for prediction of RAP by combining common clinical variables, aiming to aid clinical decision making and improve curative effect. Additionally, we externally validate the model using a separate cohort from the First People's Hospital of Foshan.

Patient Characteristics.
In this study, we reviewed the charts of patients who were diagnosed with radiotherapyinduced dysphagia after radiotherapy for NPC between January 2012 and January 2018 in Sun Yat-Sen Memorial hospital. The diagnostic criteria of dysphagia were referred to previously studies using video fluoroscopy swallowing study (VFSS) [11]. Eligibility criteria for inclusion were described as follows: (a) age ≥ 18 years, (b) history of radiotherapy for NPC, and (c) evidence of dysphagia. Excluded criteria were as follows: (a) with NPC metastasis involving the low cranial nerves leading to dysphagia, (b) with evidence of dysphagia unrelated to radiotherapy, and (c) suffering from pneumonia. The eligible patients from Sun Yat-Sen Memorial hospital were randomly classified into training cohort and internal validation cohort. To examine the generalizability of the model, we used data from patients at the First People's Hospital of Foshan between January 2012 and January 2018 as an external validation cohort with the same inclusion and exclusion criteria.

Date Collection.
Age, sex, current smoking, diet way, outcome of Kubota water drinking test, whether treated with steroid, occurrence of RAP, blood routine, low-density lipoprotein (LDL), prealbumin, albumin, high-sensitivity creactive protein (hsCRP), and erythrocyte sedimentation rate (ESR) were attained from medical records. The TNM stage, radiotherapy technique (conventional radiotherapy) or intensity-modulated radiotherapy (IMRT), chemotherapy, the maximum radiation dose of nasopharynx gross tumor volume (D max of the GTVnx), and the maximum radiation dose of lymph node gross tumor volume (D max of the GTVnd) were also recorded. All patients were restaged based on the 8th American Joint Committee on Cancer Union for International Cancer Control TNM staging manual [12]. The primary endpoint was the occurrence of RAP.
RAP was defined as the pneumonia caused by inhalation of food or vomitus [13], which was confirmed by definite findings in chest X-ray and findings of pneumonia in radiographs according to previous study [14]. The RAP was diagnosed with the following criteria [15]: (a) patients had both clinical manifestations, laboratory, and radiographic evidence of pneumonia and (b) no evidence of pneumonia caused by microorganisms.
2.3. Construction of the Nomogram. The construction of the nomogram was performed in training cohort in Renvironment. First, we used univariate Cox proportional hazard to reduce candidate predictors according to P < 0:05. Second, the independent predictors of RAP were defined by the multivariate Cox proportional hazard regression model with backward-selection procedure using Akaike's Information Criterion [16]. Thus, the nomogram was formulated by the R package of rms according to the result of multivariate analysis.

Validation and
Calibration of the Nomogram. The internal validation and external validation of the nomogram were performed by 1000 bootstrap resamples. The concordance index (C-index) was used to evaluate the discrimination ability of the nomogram. The values of the C-index ranged from 0.5 to 1.0, which means a random chance when it is 0.5, and perfect discriminate ability when it is 1 [17]. Calibration for the 1-year and 3-year pneumonia-free survival (PFS) was performed via comparing the predicted survival with the observed survival after bias correction.

Risk Group Stratification and Clinical
Usefulness of the Nomogram. Risk scores of every patient can be calculated based on the established nomogram. The maximally selected rank statistics as implemented in the "maxstat" R package was conducted to stratified patients into high-risk and lowrisk of RAP with a cutoff value of 12.3. Then, the cutoff value was applied to these two validation cohorts. Kaplan-Meier survival curves were formed compared with the log-rank test. To evaluate the clinical usefulness of this predictive model, the decision curve analysis (DCA) was performed through calculating the net benefits for a range of threshold probabilities among these three cohorts [18]. DCA was used to assess whether the decisions based on the current nomogram could improve patient's outcome.
2.6. Statistical Analyses. Continuous variables were converted to categorical variables according to the median number [19]. Proportional hazard assumption was verified by the Schönfeld test. All statistical tests were performed using the R for Windows (version 3.4.2, http://www.r-project.org/). The Cox proportional hazards regression model analysis was conducted using the "survival" package and "MASS" package. The "rms" package was used to perform the nomogram and calibrations plots. The function "stdca. R." was used to perform the DCA. All statistical tests were twotailed, and P value < 0.05 was considered significant.

Clinical Characteristics.
A total of 453 dysphagic patients with NPC were recruited from Sun Yat-Sen Memorial hospital. Randomly divided into two sets at a ratio of 2 : 1, 302 patients were assigned to training cohort and 151 patients to internal validation cohort. The external validation cohort consisted of 203 patients with dysphagia from the First People's Hospital of Foshan. The screening process was shown in Supplementary Figure 1, and the clinical characteristics of patients were listed in Table 1. The median follow-up time was 2.5 years (interquartile range (IQR) 0.7-5.4) for the training cohort, 2.5 years (IQR 1.0-6.0) for the internal validation, and 2.3 years (IQR 1.2-4.4) for the external validation cohort. The ratio of patients suffering from RAP at least once in the next 3 years after diagnosis of dysphagia was 19.5% (59/302), 21.2% (32/151), and 15.3% (31/203) in the training, internal validation, and external validation cohorts, respectively. Regarding the incidence of RAP, there was no significant difference among three groups (P = 0:312).

Predictors for RAP and Nomogram Construction.
The univariate Cox regression analysis revealed 12 clinical variable candidates in the training cohort (P < 0:05), which included diet, Kubota water drinking test grade, steroid usuage, D max of the GTVnd, traditional radiotherapy, neutrophil count, blood Hb, LDL, prealbumin and albumin, hsCRP, and ESR. Among these factors, 4 significant predictors (including Kubota water drinking test grade, D max of the GTVnd, neutrophil count, and ESR) were  Figure 2, P = 0:6889). Then, the nomogram was established based on the overall consideration of these four factors (Figure 1). The nomogram indicated that swallowing function and ESR had major contribution to RAP occurrence, followed by D max of the GTVnd and neutrophil count (Figure 1). Each factor of these variables had a corresponding score on the point scale. It was accessible to estimate the probability of RAP based on the total score with a clear boundary.

Calibration and Validation of the Nomogram.
In the training cohort, the C-index was 0.749 (95% confidence interval (CI), 0.681 to 0.817) suggesting a favorable discrimination. The calibration curves for the RAP rate at 1 year and 3 years indicated favorable agreement between the predict model and actual observation (Figures 2(a) and 2(b)). The satisfactory calculation of the nomogram was confirmed using the internal validation cohort (Figures 2(c) and 2(d)) and external validation cohort (Figures 2(e) and 2(f)). Moreover, the C-index was 0.743 (95% CI, 0.669 to 0.818) and 0.722 (95% CI, 0.606 to 0.838) in internal and external validation cohort, respectively, both with a good discrimination.

Performance of the Nomogram in Stratifying Risk of
Patients and Its Clinical Implication. By applying our nomogram, the patients were divided into low-risk and high-risk groups with a cutoff value of 12.3. Satisfactory discrimination between RAP of the high-risk and low-risk patients was observed in the training cohort (Figure 3(a), P < 0:0001), also confirmed by both internal (Figure 3(b), P = 0:0008) and external validation cohort (Figure 3(c), P = 0:005). Therefore, our nomogram served well in identifying the high-risk RAP patients after diagnosis of dysphagia. DCA was formed to estimate the usefulness of the model in a clinical context as shown in Figure 4. These plots  (Figures 4(a) and 4(b)). The usefulness of this nomogram to predict the 1-year PFS in external cohort is not satisfactory (Figure 4(e)).

Discussion
We established and validated an effective and useful predictive nomogram model to identify patients at high risk of RAP among those with dysphagia after radiotherapy for NPC. This novel prediction instrument was successfully internally and externally validated in separate cohorts and showed good discrimination and calibration. The model incorporated four factors that should be evaluated with priority in clinical practice, including Kubota water drinking test grades, D max of the GTVnd, neutrophil count, and ESR. Our data indicated that a higher Kubota water drinking test grade predicted for a higher risk of RAP. Kubota water drinking test, a simple and routine test for dysphagic patients, is used to assess the severity of dysphagia, and the higher grade indicates more serious swallowing dysfunction [11]. In line with these findings, there is emerging evidence that patients with poor swallowing function were more likely to develop RAP [20,21], and careful oral management for swallowing can reduce the incidence of RAP [22,23]. Moreover, aspiration and dysphagia after radiotherapy are regarded as the main cause of RAP in NPC patients. Identification of early swallowing dysfunction with Kubota water drinking test could screen out patients at a high-risk of RAP and facilitate targeted follow up and clinical decision making to minimize the risk of RAP. Increased D max of the GTVnd was found to increase the risk of RAP in our patients. Previous researches have found a close relationship between RAP and increased radiation dose [24,25]. D max of the GTVnd refers to the maximum radiation dose of neck lymph node gross tumor volume where the low cranial nerves lie. In addition, neck radiotherapy leads to dysphagia by damaging the neck fat, fascia, and neuron axons [26]. All the above might be the underlying mechanism for the association of increased D max of the GTVnd with risk of RAP.
Remarkably, our study found that the neutrophil count also impacted on risk of RAP. After radiotherapy, tumor cells could recruit inflammatory cells from bone-marrow, including macrophages and neutrophil [27]. Several studies have suggested that higher circulating neutrophil count indicated worse overall survival outcomes in different tumor models in patients undergoing radiotherapy [28,29]. High levels of neutrophil reflected the systemic inflammatory state [30]. Patients with high levels of neutrophil count were more susceptible to infection. Our results also found that ESR was an independent risk factor for RAP. ESR is nonspecific but one of the most commonly used laboratory markers for systemic inflammatory response in clinical practice [31]. ESR might promote RAP through enhancing postradiation inflammatory reaction.
Our study possessed several strengths. First, our nomogram was established through analysis of easily measured and routinely available predictors in a well-characterized training cohort of dysphagic patients. Second, our model has been successfully validated in two independent cohorts, thereby improving the generalizability as well as credibility. Both physicians and patients could utilize this easy-to-use model to assess the risk of RAP after the occurrence of dysphagia. Identifying patients at high risk for RAP benefits personalized treatment. However, as for those patients with dysphagia who need additional rehabilitation exercise and close follow-up, the intervention remains controversial [32]. This model would have tremendous help to address such