Modeling the risk of radiation pneumonitis in esophageal squamous cell carcinoma treated with definitive chemoradiotherapy

To develop and validate a nomogram for the prediction of symptomatic radiation pneumonitis (RP) in patients with esophageal squamous cell carcinoma (ESCC) who received definitive concurrent chemoradiotherapy. Clinical factors, dose-volume histogram parameters, and pulmonary function parameters were collected from 402 ESCC patients between 2010 and 2017, including 321 patients in the primary cohort and 81 in the validation cohort. The end-point was the occurrence of symptomatic RP (grade ≥ 2) within the first 12 months after radiotherapy. Univariate and multivariate logistic regression analyses were applied to evaluate the predictive value of each factor for RP. A prediction model was generated in the primary cohort, which was internally validated to assess its performance. In the primary cohort, 31 patients (9.7%) experienced symptomatic RP. Based on logistic regression model, patients with larger planning target volumes (PTVs) or higher lung V20 had a higher predictive risk of RP, whereas the overall risk was substantially higher for three-dimensional conformal radiotherapy (3DCRT) than intensity-modulated radiotherapy. On multivariate analysis, independent predictive factors for RP were smoking history (P = 0.035), radiotherapy modality (P < 0.001), PTV (P = 0.039), and lung V20 (P < 0.001), which were incorporated into the nomogram. The areas under the receiver operating characteristic curve of the nomogram in the primary and validation cohorts were 0.772 and 0.900, respectively, which were superior to each predictor alone. Non-smoking status, 3DCRT, lung V20 (> 27.5%), and PTV (≥ 713.0 cc) were significantly associated with a higher risk of RP. A nomogram was built with satisfactory prediction ability.


Introduction
Esophageal cancer (EC) is one of the leading causes of cancer-related death worldwide, with an estimated 508,585 deaths yearly [1]. Definitive chemoradiotherapy (CRT) is the standard care in the management of unresectable EC [2,3]. Radiation pneumonitis (RP) is one of the major complications after thoracic irradiation. Despite low mortality, RP can result in respiratory insufficiency, seriously affect patients' quality of life, influence the completion of radiotherapy and following treatment, and even reduce the curative effect, particularly for patients with symptomatic RP [4]. Although the survival benefit of concurrent CRT over radiotherapy alone is evident, the former has been demonstrated to be correlated with an increased risk of RP [5]. Therefore, decreasing the incidence of RP is of critical clinical significance for EC patients who received concurrent CRT.
Kaiqi Lan, Cheng Xu and Shiliang Liu contributed equally to this study.
An increasing number of studies have investigated predictors for RP [4][5][6][7][8][9][10][11][12]. Various risk factors, including clinical characteristics, dosimetric parameters, treatment-related variables, as well as serum biomarkers have been reported to be associated with the occurrence of RP [4][5][6][7][8][9][10][11][12]. However, the majority of these reports focused on patients with lung cancer, and data regarding EC patients undergoing definitive CRT is limited. Considering the differences in baseline pulmonary function and in the level of radiation dose between lung cancer and EC patients, it is unreasonable to extrapolate risk factors for RP in lung cancer to EC. On the other hand, despite the significantly increasing use of intensity-modulated radiotherapy (IMRT) in recent years, its impact on RP remains unclear in EC.
Therefore, the purpose of study was to investigate the associations between clinical factors, dose-volume histogram (DVH) parameters, as well as pulmonary function parameters, and the risk of symptomatic RP in patients with esophageal squamous cell carcinoma (ESCC) who received definitive concurrent CRT. Then a nomogram model predicting RP was developed and validated to guide clinical decision-making.

Patients
All consecutive ESCC patients who underwent definitive CRT at our institution from January 2010 to October 2017 were retrospectively analyzed. Inclusion criteria were defined: pathologic confirmation of stage I-IVa ESCC according to the 8th TNM staging system of the American Joint Committee on Cancer [13], receipt of concurrent CRT with curative intent, DVH data retrievable from treatment planning system, radiographic images and symptom assessments available to evaluate the occurrence of RP, and the availability of pulmonary function tests prior to CRT. Patients with prior or concomitant malignancy, those with previous thoracic radiotherapy or surgery, and those with incomplete records were excluded. A total of 402 eligible ESCC patients were included, including 321 patients in the primary cohort and 81 patients in the validation cohort based on a randomization ratio of 4:1. This study was approved by the Institutional Review Boards of Sun Yat-sen University Cancer Center. Since this was a retrospective analysis of routine data, we requested, and were granted, a waiver of individual informed consent from the ethics committee. Patient records/information was anonymized and deidentified before analysis.

Data collection
The following data were collected for each patient from the medical records: medical history, patient and tumor characteristics, treatment information, radiographic images, and symptom assessments. Pulmonary function parameters included forced expiratory volume in the first second and diffusing capacity for carbon monoxide. DVH parameters were extracted from the treatment planning system: planning target volume (PTV), total lung volume (TLV), mean lung dose (MLD), and the percentage of lung volume receiving more than x Gy (V x ), ranging from 5 to 30 Gy in increments of 5 Gy.

Treatment
A fraction of patients received one to four cycles of induction chemotherapy prior to CRT, and all patients received concurrent platinum-or taxane-based chemotherapy during radiotherapy, as determined by the multidisciplinary team. For radiotherapy, gross tumor volume (GTV) encompassed the primary tumor and involved lymph nodes. Clinical target volume (CTV) was defined as the GTV plus a 3-cm margin in proximal and distal direction and a radial margin of 0.5-1.0 cm. PTV was defined as CTV plus a 0.5-0.8 cm margin to account for setup uncertainty. Patients were treated with three-dimensional conformal radiotherapy (3DCRT) or IMRT to deliver the prescribed dose of 50-70 Gy in 25-35 fractions. All treatment plans were delivered with 6-or 8-MV photon beams. For IMRT plans, the prescribed dose was given to the PTV, and normalization was set to the PTV mean dose in the optimization and evaluation processes. Static IMRT with five to seven coplanar beams or volumetric modulated arc therapy with one arc was generated. Dose constraints for normal tissues were defined as follows: the maximum dose of spinal cord < 45 Gy; lung V 5 < 65%, V 20 < 30%, and MLD < 17 Gy; heart V 30 < 40% and mean dose < 28 Gy.

Evaluation of RP and follow-up
Patients were followed 1 month after CRT, then every 3 months during the first 2 years, every 6 months for the next 3 years, and then annually. Chest computed tomography was performed at each visit. The relevant clinical symptoms were recorded by the treating physician in medical records.
The end-point for this study was the occurrence of symptomatic RP (grade ≥ 2) within the first 12 months after radiotherapy, which was diagnosed by clinical 1 3 symptoms, chest imaging, and evidence of medication and treatment in the medical records. RP was graded based on National Cancer Institute Common Toxicity Criteria for Adverse Events version 4.0 (CTCAE v4.0).

Statistical analysis and modeling
Age, tumor length, radiation dose, and pulmonary function parameters were grouped by the median value as cut-offs. Univariate and multivariate logistic regression models were conducted to analyze possible predictors for RP. All variables in the univariate analysis were assessed in multivariate analysis (backward stepwise). Overall survival (OS) and progression-free survival (PFS) were calculated from the date of diagnosis using the Kaplan-Meier method, and differences between the groups were examined by the log-rank test.
Factors with significant predictive value in multivariate analysis were used to build the nomogram in the primary cohort. Then the nomogram was validated in the validation cohort. The performance of the nomogram was assessed by the area under the receiver operating characteristic (ROC) curve (AUC), calibration curve using 1000 bootstrap resamples, and decision curve analysis (DCA). Calibration curve was generated to compare the predicted with the observed probability of RP. DCA was employed to evaluate the clinical usefulness of the nomogram. Moreover, the optimal cut-offs of the continuous parameters in the nomogram was calculated using the ROC curves. Statistical analyses were performed using SPSS 22.0 software (SPSS Inc., Chicago, IL) and R software (version 3.4.3). A P value < 0.05 was considered to be statistically significant.

Patient characteristics
Patient and treatment characteristics of 402 ESCC patients who met the inclusion criteria are summarized in Table 1. For the primary cohort (n = 321), the median age was 59 years (range 29-73 years) and the majority of the tumors were located in the upper/middle esophagus (86.0%). Of them, 218 patients (67.9%) were current or former smokers, and COPD accounted for only 5.6%. A total of 195 patients (60.7%) received induction chemotherapy prior to concurrent CRT. The majority of patients (78.2%) were treated with IMRT, and the rest with 3DCRT, with the median radiation dose of 60.0 Gy (range 36-70 Gy) in 1.8-2.15 Gy daily fractions. A small portion of patients (5.3%) received a dose of < 50.0 Gy owing to treatment-related toxicity. The median treatment time of radiotherapy was 40 days (range 25-63 days). Owing to grade ≥ 3 hematological toxicity during radiotherapy, 23 patients (7.2%) prolonged overall treatment time of ≥ 7 days. As for DVH parameters, the median PTV was 683.1 cc (range 253.0-1566.0 cc) and median TLV was 3240.3 cc (range 1697.3-7961.9.0 cc). MLD ranged from 4.9 to 21.5 Gy, with a median of 14.8 Gy. Median V 5 , V 10 , V 15 , V 20 , V 25 , and V 30 of lung were 70.1%, 49.5%, 36.2%, 27.3%, 20.6%, and 14.8%, respectively.

RP and survival
The median follow-up time was 21.5 months (range 3.5-108.8 months) for the primary cohort. During follow-up, 159 patients (49.5%) had grade 1 RP, 27 (8.4%) had grade 2 RP, 3 (0.9%) had grade 3 RP, 1 (0.3%) had grade 4 RP, and no patients had grade 5 RP. In total, 31 patients (9.7%) experienced symptomatic RP. The median interval from the completion of radiotherapy to the diagnosis of RP was 65 days (range 20-137 days). Patients who developed symptomatic RP had a trend of worse OS but similar PFS compared with those who did not, with corresponding 1-year OS rates of 72.1% and 61.3% (P = 0.055; Supplementary Fig. 1).

Univariate analysis
The comparisons of clinical and dosimetric factors between patients with or without symptomatic RP are listed in Table 2 and Supplementary Fig. 2A. For clinical characteristics, univariate analysis showed that age, smoking history, performance status, primary tumor length, and radiotherapy modality were correlated with the development of RP. Among dosimetric parameters, PTV, Lung V 5 to V 30 , and MLD were significantly associated with the risk of RP (P < 0.05 for all). Nevertheless, no pulmonary function parameters were significant risk factors for RP.

Multivariate analysis
On multivariate analysis, smoking history [odds ratio (OR) 3.445, P = 0.035], radiotherapy modality (OR 11.003, P < 0.001), PTV (OR 1.003, P = 0.039), and lung V 20 (OR 1.396, P < 0.001) were independent predictive factors for symptomatic RP (Table 3). Among these factors, smoking history was found to protect against RP. Figure 1 shows the predictive probabilities of symptomatic RP as a function of PTV or lung V 20 for the two radiation modalities based on logistic regression model. Patients with larger PTVs or higher lung V 20 had a higher predictive risk of RP, whereas the overall risk was substantially higher for 3DCRT than IMRT. For example, when PTV is 630.0 cc, the probability of developing symptomatic RP would be 5.0% in patients treated with IMRT versus 20.2% in those with 3DCRT.

Nomogram development and validation
Based on the multivariate analysis, a nomogram model was built to predict the risk of symptomatic RP, including smoking history, radiotherapy modality, PTV, and lung V 20 (Fig. 2).  (Fig. 3A). The optimal cut-offs for PTV and lung V 20 were 713.0 cc and 27.5%, respectively. Additionally, the calibration curve showed favorable agreement between the prediction by nomogram and the actual observation (Fig. 3B). Also, the DCA exhibited satisfactory positive net benefits of the model among the majority of threshold probabilities, indicating excellent clinical utility (Fig. 3C).
In the validation cohort (n = 81), 37 (46.3%) patients had grade 1 RP, 8 (9.9%) had grade 2 RP, 2 (2.5%) had grade 3 RP, and no patients had grade 4 or 5 RP. A total of ten patients (12.3%) experienced symptomatic RP. The comparisons of lung dosimetric factors between patients with or without symptomatic RP are listed in Supplementary Fig. 2B. The 1-year OS rates for patients with or without symptomatic RP were 70.0% and 73.4%, respectively (P = 0.607; Supplementary Fig. 1). Application of the nomogram model in the validation cohort yielded an excellent AUC of 0.900 (95% CI 0.779-1.000) (Supplementary Table 2 and Supplementary Fig. 3).

Discussion
Symptomatic RP is a common side effect of radiotherapy for EC with incidence of 5.7-35.0% [5][6][7][8][9][10][11], which is confirmed by our study. In this cohort of ESCC patients undergoing definitive CRT, smoking history, radiotherapy modality, PTV, and lung V 20 were significant predictive factors for symptomatic RP. More importantly, a nomogram has been built and validated, indicating satisfactory prediction ability. Thus, this prediction model could help clinicians select high-risk patients who may benefit from modified treatment approaches to reduce the risk of RP prior to the initiation of treatment.
In recent years, radiation techniques have evolved from 3DCRT to IMRT and proton therapy in EC. Numerous dosimetric studies have well demonstrated the superiority of IMRT over 3DCRT in improving target coverage and sparing adjacent organs, but whether the dosimetric benefits could translate into clinical benefit, especially reducing the incidence of radiation-related toxicities, remains inconclusive due to the lack of prospective evidence [14]. In a large-scale retrospective study reported by He et al., IMRT significantly reduced the incidence and postponed the onset of pleural effusion in EC patients, compared to 3DCRT [15]. However, Haefner et al. found no evident difference in acute toxicities between IMRT and 3DCRT, possibly due to the small sample size of the cohort and the higher radiation dose in the IMRT group [16]. In our study, after adjusting for smoking history, PTV, and Lung V 20 , IMRT was associated with a substantially lower risk of RP than 3DCRT. This is in consistence with the secondary analysis of results from RTOG 0617, which prospectively demonstrated that IMRT group had significantly less severe RP than 3DCRT group in locally advanced non-small cell lung cancer (3.5% vs. 7.9%, P = 0.039) [17]. In addition, IMRT could also reduce the incidence of postoperative pulmonary and cardiac complication in EC patients who received neoadjuvant CRT and surgery, as reported by Lin et al. [18]. Furthermore, IMRT was associated with more favorable survival outcomes than 3DCRT in EC [14]. Collectively, despite the paucity of prospective evidence, the current findings strongly suggest the routine use of IMRT in EC.
As a novel radiation technique with superior physical properties, proton therapy has the potential to improve normal tissue sparing as compared to 3DCRT or IMRT [14]. COPD, chronic obstructive pulmonary disease, ECOG Eastern Cooperative Oncology Group, 3DCRT three-dimensional conformal radiation therapy, IMRT intensity-modulated radiation therapy, V x percentage of the total lung volume receiving more than x Gy, IQR interquartile range, MLD mean lung dose, PTV planning target volume, TLV total lung volume, FEV1 forced expiratory volume in the first second, DLCO diffusing capacity for carbon monoxide a Single agent chemotherapy during radiotherapy (taxane or platinum)  RP radiation pneumonitis, CI confidence interval, COPD chronic obstructive pulmonary disease, ECOG Eastern Cooperative Oncology Group, 3DCRT three-dimensional conformal radiation therapy, IMRT intensity-modulated radiation therapy, V x percentage of the total lung volume receiving more than x Gy, MLD mean lung dose, PTV planning target volume, TLV total lung volume, FEV1 forced expiratory volume in the first second, DLCO diffusing capacity for carbon monoxide a Single agent chemotherapy during radiotherapy (taxane or platinum) Fig. 1 Model to represent the estimated risk of symptomatic radiation pneumonitis as a function of planning target volume (a) and lung V 20 (b) for the two radiation modalities based on logistic regression model in the primary cohort. Red line represents three-dimensional conformal radiotherapy (3DCRT) and blue line represents intensitymodulated radiotherapy (IMRT). Shaded regions represent 95% point-wise confidence intervals Several studies have investigated the clinical advantages of proton therapy compared to photon therapy. For EC patients who underwent neoadjuvant CRT, proton therapy was superior to IMRT in reducing incidence of pulmonary complications [18]. Recently, Lin et al. reported that the proton arm experienced numerically fewer cardiopulmonary toxicities compared with IMRT arm in a phase IIB randomized trial for EC [19]. The ongoing larger cooperative group studies will clarify the clinical benefit of proton therapy in EC, such as NRG-GI006. There has been a general consensus that DVH parameters are important predictors for RP. However, there is still no recommendation of dose-volume constraints for EC. Wang et al. reported that the volume of the lung spared from doses of ≥ 5 Gy was the only independent dosimetric factor associated with pulmonary complications [20]. Cho  [21]. Likewise, a recent study demonstrated the strong correlation between MLD and severe RP in 416 EC patients undergoing CRT [22]. Consistent with studies reported by Asakura et al. and Shaikh et al. [9,23], all DVH parameters (lung V 5 -V 30 and MLD) were significantly associated with RP in univariate analysis in our cohort. Of them, V 20 was the only independent predictor in multivariate analysis, with the optimal threshold value of V 20 (27.5%). According to the previously published literature, V 20 can be affected by various factors, including radiotherapy modality, treatment volume, different priorities in normal organ constraints, or beam arrangements. In our study, logistic regression in analysis of interactions was performed and indicating that no potential interactions existing between PTV, radiotherapy modality and V 20 . The failure of the other dosimetric parameters to be retained significance in the multivariate model could be explained by their potential correlation with V 20 . Thus, other dosimetric parameters should also be taken into consideration when performing treatment planning for EC.
In addition to lung dosimetric parameters, we observed that patients with greater PTVs had a remarkably higher incidence of RP. In line with our results, Cui et al. also reported the strong correlation between PTV and the occurrence of RP in elderly EC [24]. Considering PTV is a variable that could be modified, smaller radiation volumes might reduce the risk of RP. A meta-analysis reported that neither local control rates nor survival outcomes differed significantly between elective nodal irradiation and involved-field irradiation in ESCC, whereas incidences of severe RP and radiation esophagitis were significantly lower in the latter group [25]. Therefore, involved-field irradiation should be considered in clinical practice, especially for elderly patients.
Among clinical factors, smoking history was also found to be a protective factor for RP in the current study, which is consistent with previous studies [26,27]. In a meta-analysis of clinical risk factors on the incidence of symptomatic RP, smoking was identified to protect against RP (OR 0.6, P = 0.008) [28]. The possible explanations are smokingassociated hypoxia and a decreased inflammatory reaction induced by irradiation among smokers [29,30], but the underlying mechanism for the smoking effect on RP remains unclear. Of note, this information should not encourage tobacco consumption in cancer patients.
Compared with each separate predictor, integrating predictive factors to develop a statistical model will further improve predictive accuracy. So far, only one study reported by Wang et al. had built a combined model to predict severe RP in EC [22]. However, this study included a fraction of patients without concurrent chemotherapy during radiotherapy, which might affect the incidence of RP. Moreover, this study incorporated not only pretreatment factors but also the changes of inflammatory indexes during radiotherapy into the nomogram [22]. Despite excellent discriminatory power of the model, it is of less value to guide decision-making before the initiation of treatment.
The association between RP and prognosis has been well established in lung cancer. Inoue et al. reported that 3-year survival rates of patients with lung cancer who experienced no, mild, and severe RP were 33.4%, 38.2%, and 0%, respectively, and severe RP was the most important factor determining poor survival [31]. Another study also demonstrated that severe radiation pneumopathy was an independent prognostic factor for survival in lung cancer [32]. Compared with lung cancer, there is currently insufficient data regarding the association between RP and survival in EC patients. Recently a retrospective study from the University of Texas MD Anderson Cancer Center found that pulmonary complication was associated with worse survival (5-year OS rates 23.1% with vs. 47.4% without, P < 0.001) [33]. Consistently, the current study showed a trend of worse OS in patients with symptomatic RP in the primary cohort. However, OS rates were comparable between patients with or without RP in the validation cohort, which could be explained by the small sample size.
It should be noted that our work also has several limitations. Firstly, selection bias existed due to the retrospective nature of this study. Secondly, the sample size of the validation set is relatively small, and an independent external validation is needed. Thirdly, although all patients included were treated with concurrent CRT, treatment modalities such as the utility of induction chemotherapy and chemotherapy regimens were not identical in this study. However, these factors were not correlated with the risk of RP in univariate analysis. Finally, serum biomarkers as well as radiomics-based features were not incorporated into the study, which might further promote predictive ability of the model.
In conclusion, non-smoking status, 3DCRT, lung V 20 (> 27.5%), and PTV (≥ 713.0 cc) were significantly associated with a higher risk of symptomatic RP in ESCC patients treated with definitive CRT. Then a nomogram was built and validated, exhibiting satisfactory prediction ability. Further studies are warranted to verify the efficacy of the predicting model and to explore potential strategies for minimizing the risk of RP in high-risk patients.