Clinical-Radiomics Nomogram for Risk Prediction of Esophageal Fistula in Patients with Esophageal Squamous Cell Carcinoma Treated by IMRT or VMAT

Background and purpose: To analyze the predictive factors and establish a prediction model of esophageal �stula (EF) in patients with esophageal squamous cell carcinoma (ESCC) received intensity-modulated radiotherapy (IMRT) or Volumetric Modulated Arc Therapy (VMAT). Materials and Methods: Patients with ESCC treated with IMRT or VMAT from 2013 to 2020 in Xijing hospital were retrospectively analyzed. 43 patients with EF and 129 patients without EF were included in the analysis by 1:3 propensity score matching (time of diagnosis, gender). The clinical characters and radiomic features were recorded and extracted. Univariate and multivariate stepwise logistic regression analyses were provided to determine the risk factors associated with EF. Results: The median follow-up time was 23.96 months (range 1.3-104.9 m), and the median OS of EF patients was 13.1 months. 1158 radiomics features were extracted and 8 radiomics features were selected. The area under the receiver operating characteristic curve (AUC) value of radiomic signature calculated by selected features for predicting EF was 0.794. Multivariate analysis showed that the tumor length, tumor volume, T stage, lymphocyte rate and grade 4 esophagus stenosis were related to EF, and the AUC value of clinical nomogram for predicting EF was 0.849. The clinical-radiomics nomogram had the best performance in predicting EF with an AUC value of 0.896. Conclusions: The clinical-radiomics nomogram


Background
Esophageal cancer (EC) is the 6th most common cancer and ranks 5th in all tumor related death in China [1].About 90% EC in China are esophageal squamous cell carcinoma (ESCC) [2].Radiotherapy with or without chemotherapy is the current standard treatment for unresectable locally advanced EC [3].
Previous studies have established EF prediction models using T4, N3, maximum thickness of the tumor, ulcerative type, stenosis, low Body Mass Index (BMI), etc [7][8][9][10][11][12][13][14][15][16][17].In addition to clinical factors, quantitative imaging features might provide promising information for risk prediction.Radiomics, a technology to e ciently mine quantitative image features from standard medical imaging, has been widely used in terms of prognosis prediction[18] [19].Previous studies and ours have con rmed that Computed Tomography (CT)-based radiomics can predict overall survival and progression-free survival in patients with esophageal cancer, but the predictive value in EF is less reported [20] [21].In two studies reporting the predictive value of radiomics in EF, not all EC patients received radiotherapy and the prediction model for EF was not established [22] [23].Furthermore, few studies have reported the risk factors for EF in patients with ESCC treated by intensity-modulated radiotherapy (IMRT) or Volumetric Modulated Arc Therapy (VMAT) [11].To the best of our knowledge, no studies have reported the clinical-radiomics nomogram for risk prediction of EF in ESCC patients treated by IMRT or VMAT.
This study was aimed to analyze the predictive factors of EF in ESCC patients received IMRT/VMAT, and further establish a prediction model for risk of EF based on clinical factors and radiomics features.

Patients
Patients with ESCC treated by IMRT or VMAT at our institution from January 2013 and December 2020 were collected.EF was diagnosed by esophageal barium meal radiography, CT, MR or endoscopy.In this study, we only focused on EF which occurs during or after radiotherapy.The inclusion criteria and exclusion criteria were described in Supplementary A1.This retrospective study was approved by the local ethics committee (KT20172035-1).

Clinical Data Collection
All data were collected from EMR and radiotherapy system.The following clinical characters, hematological parameters and treatment parameters were recorded and analyzed.Basic clinical characters include age at diagnosis, gender, PS score, smoking history, drinking history and BMI.Tumor characteristics include the location of tumor, T stage, N stage, tumor type, stenosis grade, the longitudinal length of tumor, tumor axial maximum thickness and tumor volume.Treatment characteristics include induction chemotherapy, concurrent chemotherapy, chemotherapy regimen, IMRT or VMAT, total radiation dose, single radiation dose, and radiation elds.Hematological (laboratory) parameters include leukocyte count, neutrophils, lymphocytes, lymphocyte ratio (LR), hemoglobin, platelets, and serum albumin.All hematological parameters were collected within 1 week before radiotherapy, and the neutrophil/lymphocyte ratio (NLR), hemoglobin/lymphocyte ratio (HLR) and platelet/lymphocyte ratio (PLR) were calculated according to the collected data.Onodera prognostic nutritional index (OPNI or PNI) was calculated as follows: PNI = serum albumin (g/L) + 5× total number of peripheral blood lymphocytes (×10 9 /L).Tumor staging was performed according to the 7th edition of the American Joint Committee on Cancer (AJCC) staging criteria.The type of EC was determined on the basis of imaging ndings on barium meal, which was determined by less than 2 senior radiologists.The stenosis grade was determined according to the ratio of the diameter of the stenosis segment and the normal segment of the esophagus on the barium meal image, as Gui et al. reported [24].In this study, we used their grading method and divided the esophageal stenosis into 4 grades: grade I, 0%-24%; grade II, 25%-49%; grade III, 50%-74%; grade IV, 75%-100%.Overall survival (OS) was calculated from the start of radiotherapy to death or follow-up date.

Image Acquisition and Tumor segmentation
The ESCC patients underwent a standard chest contrast-enhanced CT scanning before radiotherapy with Brilliance Big Bore (Philips Healthcare, Cleveland OH, USA) scanner according to the following acquisition parameters: 120 kV tube voltage, 300 mA tube current, 5 mm slice thickness, 512×512 pixels in-plane resolution, Standard B (body) reconstruction kernels.The tumor was delineated as the region of interest (ROI) by one radiologist with ve years of clinical-diagnosing experience via the ITK-SNAP software (www.itksnap.org).The other two radiologists with 10 years of experience corrected the ROIs by consensus.

Feature extraction, selection and radiomic signature construction
The Z-score normalization, which could reduce the variability between images from different patients, was adopted for image of each patient.The 1223 radiomic features were extracted by Pyradiomics (version 2.2.0) [25].To reduce any type of bias or over-tting caused by too many features, feature selection was conducted by least absolute shrinkage and selection operator (LASSO) [26] [27].The details of feature extraction and selection were shown in Supplementary A2.Radiomic signature was developed as radiomics scores (radscore) calculated by a linear combination of the selected features that were weighted by their respective coe cients [28].To verify the association of radscore with patient survival time, patients were classi ed into low-risk group and high-risk group according to the median of radscore and plot survival curves.

Establishment of clinical-radiomics nomogram
The clinical nomogram was constructed by the useful clinical risk factors selected from all clinical parameters mentioned above.To demonstrate the incremental value of the radiomic signature to the clinical risk factors for individualized prediction of EF, the clinical-radiomics model was established by multivariate logistic regression analysis using useful clinical risk factors and radiomic signature.To provide the clinician with a quantitative tool for individualized assessment of the LRFS, we built a nomogram on the basis of the clinical-radiomics model.The nomogram calibration curve was assessed by plotting the actual probabilities against the nomogram-predicted probabilities.Decision curve analysis was conducted to determine the clinical usefulness of the clinical-radiomics model by quantifying the net bene ts at different threshold probabilities.

Statistical analysis
Univariate and multivariate analyses were carried out using logistic regression to estimate the odds ratio (OR) and 95% con dence intervals (CIs).The association of various factors with the risk of EF were assessed by univariate logistic regression, and the best cutoff values to predict EF risk were determined using receiver operating characteristic (ROC) curves.Based on a P value < 0.1 in univariate analyses, factors were selected into the multivariate stepwise logistic regression.The nomogram for the prediction of probability of radiotherapy-related EF was established with the results of multivariate analysis.
The discrimination of the model was assessed by area under the receiver operating characteristic (ROC) curve (AUC), accuracy (ACC), true positive rate (TPR) or sensitivity (SENS), true negative rate (TNR) or speci city (SPEC), positive predictive value (PPV), negative predictive value (NPV).Calibration curve and Brier score were adopted to evaluate the nomogram.Decision curve analysis was conducted to estimate the clinical usefulness of the models.The Kaplan Meier analysis was used to plot survival curves, and log rank test was used to compare survival differences.Statistical analysis was performed using SPSS 26.0 and R version 3.3.3(www.R-project.org) for Windows.

Results
Patients' characteristics A total of 483 ESCC patients received IMRT or VMAT in our department from January 2013 to December 2020 were included.According to the time of EC diagnosis and gender, 129 patients without EF and 43 patients with EF were matched with a 1:3 ratio.The characteristics of patients were listed in Table 1 and described in Supplementary A3.

Clinical outcome
Median follow-up time was 23.96 months (range 1.3-104.9m).At the end of the follow-up, 5 patients with EF were alive and the 1-year, 2-year, and 3-year survival rates were 51.2%, 23.0%, and 14.4%, respectively.64 patients without EF were still alive, and the 1-year, 2-year, and 3-year survival rates were 80.6%, 64.7%, and 54.7%, respectively.The median OS of the 172 patients included in the analysis was 27.8 months (range 1.3-104.9m),and the median OS of EF patients was lower than that of non-EF patients with 13.1 months (range 1.3-65.3m)and 49.4 months (range 4.0-104.9m,Figure 1A, p<0.001), respectively.17 patients (39.5%) occurred EF during radiotherapy, and 26 patients occurred EF after radiotherapy.The survival of patients with EF during radiotherapy was signi cantly worse than that of patients with EF after radiotherapy, with the median OS of 8.1 months and 17.6 months, respectively (Figure 1B, p=0.041).

Performance of clinical nomogram
In the univariate analysis of the clinical variables, the longitudinal length of tumor, tumor axial maximum thickness, tumor volume, T4 stage, N2-3 stage, PNI, LR, NLR, HLR, and PLR were identi ed as signi cant factors for EF.All factors with p < 0.1 were further included in the multivariable stepwise logistic regression analysis.Eventually, the longitudinal length of tumor (≤ 8.4 cm), tumor volume (> 59.6 cm 3 ), T4 stage, grade 4 stenosis and LR (≤ 0.234) were risk factors included in the clinical model (Table 2).The performance of clinical model based on 100 times of 5-ford cross-validation was shown in Table 3.The AUC and accuracy (ACC) were 0.849 and 85.45%, respectively.The clinical nomogram based on clinical model was shown in Figure A1A.The calibration curve in Figure A1B showed good conformity between predicted and actual probability for EF.The Brier score of the clinical nomogram was 0.106 (Figure A1C).The Brier score is a measure of calibration of probabilistic prediction on a set of probabilistic predictions.
The value range is between 0 and 1, and the smaller the Brier Score, the higher the accuracy of the model.

Performance of radiomic signature
Eight radiomic features that were most useful to predict EF were selected by LASSO for calculating the radiomic signature (Figure A2, Table A2).The distribution of radiomic signature for each patient was shown in Figure A2.The performance of radiomic signature based on 100 times of 5-ford crossvalidation was shown in Table 3.The AUC and ACC were 0.794 and 81.40%, respectively.The median of radscore (-1.245) was used to divide patients into high-and low-risk groups with different OS.The lowrisk group (Radsocre-) with lower radscore had signi cantly better OS than high-risk group (Radscore+) (median OS: 46.6m vs 17.6m, p < 0.05; Figure 1C).

Performance of clinical-radiomics nomogram
Based on multivariate logistic regression analysis, the clinical-radiomics nomogram was established based on the combination of clinical factors and radiomic signature mentioned above (Figure 2A).The performance of clinical-radiomics model based on 100 times of 5-ford cross-validation was shown in Table 3.The AUC and ACC were 0.896 and 81.40% respectively, which were higher than the clinical nomogram and radiomic signature.The ROC curves were shown in Figure 3A.As shown in Figure 2B, the calibration plot showed good conformity between predicted and actual probability for EF.The Brier score of the clinical-radiomics nomogram was 0.088, which was much closer to 0, as compared with clinical model (0.106), indicating better predictive ability (Figure 2C).Finally, we performed a decision curve analysis (DCA) to evaluate the clinical utility of the models and their effective threshold ranged from approximately 10% to 75% (Figure 3C), showing that using these models were more effective than the "treat-all" or the "treat-none" strategy and the clinical-radiomics nomogram was most effective in predicting EF when the prediction probability was within this range.

Discussion
In this study, we investigated the predictive factors of EF in patients with ESCC received IMRT or VMAT based on clinical factors and radiomics features, and found that the longitudinal length of tumor, tumor volume, T4 stage, grade 4 stenosis and LR was related to the occurrence of EF.Moreover, the radiomic signature could separate patients into high-and low-risk groups in terms of different OS rate.We further established a clinical-radiomics nomogram, which performed better than clinical nomogram and radiomic signature alone.
EF is one of the serious adverse events of esophageal cancer patients during/after radiotherapy.The risk prediction model of EF in EC had been established reported by several studies [9] [24], in which only clinical parameters were considered and enrolled.Previous studies have shown the potential of CT radiomics for predicting the risk of EF and the prognosis of EC patients.Xu et.al developed a deep learning model to integrate CT imaging features and clinical factors for predicting EF caused by tumor itself and treatment in EC patients [23], which achieved a C-index of 0.901 in validation set.The study of Chao Zhu et.al constructed nomograms incorporating independent clinical risk factors and radiomic signature to predict the prognosis of malignant EF [22].Due to the complexity and poor interpretability of deep learning algorithm and its application, a simple nomogram model may be more suitable for clinical use.In this study, we developed a predictive model for EF in ESCC patients received IMRT or VMAT using CT-based radiomics and more detailed clinical parameters.Compared with the model using clinical parameters or radiomics parameters alone, the combination model has good prediction performance and accuracy with an AUC value of 0.896.The results also showed that radiomics can provide reference value for the survival of patients with EF.To the best of our knowledge, this is the rst study to establish a clinicalradiomics model for predicting EF in ESCC patients treated by IMRT or VMAT.
In this study, we found that patients who developed EF during radiotherapy had worse survival than those who developed EF after radiotherapy.This may be related to the deterioration of nutritional status caused by stula during radiotherapy or tumor progression during radiotherapy.Several studies have investigated potential factors associated with radiotherapy-related EF, including tumor and patients related factors [4] [9][10][13] [24].Pao et al. also showed baseline T4 and esophageal stenosis were EF risk factors for ESCC treated by de nitive concurrent chemoradiotherapy with IMRT [11].Our study also showed that T4 was highly predictive of EF for patients received radiotherapy.It's likely that T4 stage is a tumor that directly invades surrounding normal tissues and organs, repair of normal tissues could not compensate the shrinkage of tumor during radiotherapy and this results in the EF.Our results suggested that tumor volume, the longitudinal length of tumor, and LR also play important roles in the formation of EF.These were similar to the risk factors of developing EF investigated in patients with conventional radiotherapy and three-dimensional conformal radiotherapy [32].The correlation between modern radiotherapy technique and the occurrence of EF needs further validation.
There is no consensus on the classi cation and de nition of esophageal stenosis.Kawakami et al. [7] and Pao et al. [11] de ned esophageal stenosis as the impossibility to pass through the lesion by an endoscope, and Hu et al. [14] determined the grade of esophageal stenosis by measuring the internal diameter at the local stricture segment on a barium meal, but these methods were either complicated during manipulation or were not reproducibly validated.Gui et al. [24] classi ed the quanti cation of esophageal stenosis into 4 grades, and their study showed that not all esophageal stenosis were associated with EF formation.Only the grade 4 esophageal stenosis was a high-risk factor for EF formation [24], which was consistent with our ndings.It was hypothesized that the internal pressure caused by the tumor lead to the esophageal stenosis but the underlying mechanism of EF still remains unclear.Taniyama et al.[8] and Hu et al. [14] reported the axis of a tumor thickness on CT images was associated with EF formation.However, our study showed that the longitudinal length of tumor and tumor volume may be more susceptible factors for EF than tumor thickness.Larger tumor size indicates greater local tumor burden, which may cause more severe local normal tissue damage.Excessive tumor length may cause the reduction of lymphoid immune cells and in ammatory cells around the tumor, so as to reduce the repair function of local normal cells.Previous studies showed that ulcer type was related to the formation of EF [10][13] [14][16].However, it did not show any signi cance in our results, which might be related to the fact that our ulcer determination was based only on barium meal examination and this resulted the data bias.Some studies showed that high PLR was associated with the occurrence of EF [24] [29].Our results showed that PLR was related to the formation of esophageal stula by univariate analysis, but it did not show signi cance by multivariate analyses analysis.The value of PLR in EF formation and the determination of threshold still need further data validation.Wang et al. showed that absolute lymphocyte count (ALC) was associated with the occurrence of EF [13].However, ALC was not associated with EF in our results, and EF was more likely to occur when LR was lower than 0.234.The decrease of LR may represent the decrease of the patient's immune function, which may reveal the decline of the patient's normal tissue repair function, thus resulting in a higher incidence of EF.Xue et al. reported that PNI was an independent healing factor for EC, and the survival of EC patients with low PNI was worse [30], but our study showed that PNI had no clear value in predicting EF.Some studies have shown that the continuous improvement of nutritional status such as NRS during treatment is signi cantly related to the low incidence of EF [31][32].Due to the lack of historical data, the changes of nutritional status and indicators during radiotherapy for EF were not analyzed in this study.The nutritional score and nutritional improvement of patients with EC may have a clear relationship with the formation of EF, but further research is still needed to verify it.This study has several limitations.First, it is a retrospective study from one institution, and the small sample size may cause some selective bias.Second, the clinical-radiomics nomogram model has not been veri ed by external data, and further validation in multicenter is needed to determine its clinical applicability.

Conclusions
In this study, we developed a clinical-radiomics model for predicting EF in ESCC.The nomogram is helpful to select high-risk patients for individualized treatment.

Figures
Figures

Figure 1 The
Figure 1

Table 1 .
Characteristics of patients.

Table 3 .
Performance of clinical model, radiomic signature, and clinical-radiomics model based on 100 times of 5-ford cross-validation (mean(standard deviation)).