A Nomogram for the Determination of the Necessity of Concurrent Chemotherapy in Patients With Stage II–IVa Nasopharyngeal Carcinoma

Background The efficiency of concurrent chemotherapy (CC) remains controversial for stage II–IVa nasopharyngeal carcinoma (NPC) patients treated with induction chemotherapy (IC) followed by intensity-modulated radiotherapy (IMRT). Therefore, we aimed to propose a nomogram to identify patients who would benefit from CC. Methods A total of 434 NPC patients (stage II–IVa) treated with IC followed by IMRT between January 2010 and December 2015 were included. There were 808 dosimetric parameters extracted by the in-house script for each patient. A dosimetric signature was developed with the least absolute shrinkage and selection operator algorithm. A nomogram was built by incorporating clinical factors and dosimetric signature using Cox regression to predict recurrence-free survival (RFS). The C-index was used to evaluate the performance of the nomogram. The patients were stratified into low- and high-risk recurrence according to the optimal cutoff of risk score. Results The nomogram incorporating age, TNM stage, and dosimetric signature yielded a C-index of 0.719 (95% confidence interval, 0.658–0.78). In the low-risk group, CC was associated with a 9.4% increase of 5-year locoregional RFS and an 8.8% increase of 5-year overall survival (OS), whereas it was not significantly associated with an improvement of locoregional RFS (LRFS) and OS in the high-risk group. However, in the high-risk group, patients could benefit from adjuvant chemotherapy (AC) by improving 33.6% of the 5-year LRFS. Conclusions The nomogram performed an individualized risk quantification of RFS in patients with stage II–IVa NPC treated with IC followed by IMRT. Patients with low risk could benefit from CC, whereas patients with high risk may require additional AC.


INTRODUCTION
The cases of newly diagnosed nasopharyngeal carcinoma (NPC) reached about 129,000 in 2018 (1). More than 70% of NPC patients are in the east and southeast Asia. NPC is relatively sensitive to radiotherapy; radiotherapy is the mainstay treatment modality for nonmetastatic disease. As for stage II-IVa NPC, concurrent chemotherapy (CC) combined with radiotherapy is used to enhance radiosensitivity and local control. In the two-dimensional (2D) radiotherapy era, concurrent chemoradiotherapy (CCRT) is superior to radiotherapy in patients with stage III-IVa disease for progression-free survival (PFS) and overall survival (OS) (2). Stage II disease could also benefit from CCRT in terms of PFS and OS compared with radiotherapy alone in the 5-and 10-year follow-up (3,4). Therefore, CCRT is widely accepted for treating stage II-IVa NPC (5).
Over time, radiotherapy modality has progressed from conventional 2D radiotherapy to 3D conformal radiotherapy and then to intensity-modulated radiotherapy (IMRT) for NPC. Recently, evidence from retrospective studies has shown that induction chemotherapy (IC) plus IMRT was superior to IC plus CCRT, with comparable survival outcomes and fewer acute toxic effects in treating stage II-IVa NPC (6). Notably, CCRT has a potential risk to increase the incidence of grade 3/4 acute toxicities (2,6) and radiotherapy time, which adversely affect the survival of NPC (7). Therefore, it is critical to determine whether patients treated with IC followed by IMRT can benefit from CC. This can greatly enhance clinical decision-making.
IMRT is an advanced type of radiation therapy to conform the radiation dose to the tumor and to avoid the exposure of organs at risk (OARs) to enhance the therapeutic ratio. However, it is still challenging because of the anatomical proximity of gross tumor to OARs (8). Therefore, dosimetric inadequacy of target volumes that resulted from prescription constraints on critical OARs has been one of the most crucial independent prognosticators of survival outcomes of NPC (8)(9)(10). Dosimetric factors are commonly used to quantify the prescription isodose line coverage to the corresponding target volume and utilized to evaluate the plan quality among various radiotherapy approaches (11). It may improve patient care, too, by incorporating big data of dosimetric factors into clinical practice based on dose-volume histogram metrics of treatment plan (12,13). Recently, an international guideline in radiotherapy planning for NPC highlights dose criteria (8). In addition, dosimetric factors are regarded as score criteria for radiotherapy plan comparison in the International Radiotherapy Plan Competition (https://radiationknowledge.org/). Therefore, precision medicine needs to quantify clinical treatment plans based on dosimetric factors.
In this study, we aimed at developing a prognostic nomogram based on IMRT dosimetric signature as a risk quantification model and further identifying patient subsets who can benefit from CC in patients with stage II-IVa NPC treated with IC followed by IMRT. Figure 1 shows the workflow of this study, including (1) radiotherapy planning data restored, (2) dosimetric parameter extraction by in-house script, (3) dosimetric signature building, and (4) nomogram development and risk stratification for concurrent chemotherapy. nature of this study. A dataset of 523 patients with stage II-IVa NPC (AJCC/UICC 8th edition) was reviewed from January 2010 to December 2015. The inclusion criteria were as follows:

Study Workflow
(1) patients with biopsy-confirmed primary NPC, (2) patients received cisplatin-based IC and radical IMRT, (3) patients received no previous treatments due to other cancers, and (4) patients with restored radiotherapy planning. The exclusion criteria were as follows: (1) failed to restore the radiotherapy planning (n = 14), (2) no available radiotherapy data in mosaiq system (n = 3), and (3) radiotherapy planning could not be extracted by the in-house script (n = 16). Supplementary Figure  S1 presented the patient inclusion flowchart. Finally, 434 patients were included, of which 247 patients (IC ± CC cohort) were treated with IC + CCRT or IC + IMRT and 187 (IC ± CC + AC cohort) were treated with IC + CCRT + AC or IC + IMRT + AC.
The characteristics of the patients were collected from the hospital information systems. The chemotherapy and radiotherapy regimens are described in Supplementary S1. The primary endpoints included RFS, locoregional RFS (LRFS), distant metastasis-free survival (DMFS), and OS. RFS was defined as the time from treatment until the diagnosis of a recurrent disease. LRFS was defined as the time from treatment to local or regional recurrence. DMFS was defined as the time from treatment to data of distant recurrence. OS was defined as the date from treatment to death from any cause.

Dosimetric Parameter Extraction and Dosimetric Signature Building
The radiotherapy planning data were restored from the mosaiq system (Elekta Medical Systems, Sunnyvale, CA, USA) and were transferred to the treatment planning system Raystation. A total of 808 dosimetric parameters were calculated from the target volume for each patient through our in-house script (14), including D1-100, D1 cc-5 cc, D1min cc-5min cc. Dv% was described as the minimal dose delivered into v% of the target volume. Dv cc and Dvmin cc are the maximal and minimal dose exposed in the absolute target volume v cc. Gross tumor volume (GTV) contains the gross tumor volume of the nasopharynx (GTVnx) and gross tumor volume of regional lymph node (GTVnd). The planning GTVnx/nd (PGTVnx/nd) was obtained by expanding the corresponding GTVnx/nd with an expanded margin of 3 mm. GTVnx, GTVnd, PGTVnx, and PGTVnd were the target volumes. The least absolute shrinkage and selection operator (LASSO) algorithm was applied to select the dosimetric parameters that associated with RFS among the above-mentioned 808 features. The selected parameters were weighted by their respective coefficients and were calculated as a linear combination to construct the planning score for each patient.

Development and Validation of Nomogram
Univariate and multivariate Cox proportional hazards regression model was applied to achieve the candidate variables. A nomogram predicting RFS was constructed based on the candidate variables of 247 patients from the IC ± CC cohort and was internally validated by performing 1,000 bootstrap resamples. The concordance index (C index) was utilized to evaluate the predictive performance of the nomogram. In addition, a calibration curve was plotted to compare the actual probabilities against the prediction probabilities of the nomogram.

Classification of Patients Into Low-or High-Risk Subgroup
We computed the nomogram risk score for each patient by linear predictor function "predict" in R, which was applied for prediction based on the results of Cox model fitting functions (14). The cutoff value of the nomogram risk score was calculated by the X-tile software (version 3.6.1; Yale University, New Haven, CT, USA), which produced the largest c² value in the Mantel-Cox test (15). The 247 patients were divided into low-or high-risk group according to the optimal cutoff risk score, and the same classification criterion was performed on the 187 patients.

Statistical Analysis
All statistical analyses were performed on R software (version 3.5.2), MedCalc (version 15.6), and SPSS (version 22.0). The ttest was used for comparison between continuous variables, and chi-square test was used for comparison between categorical variables. All dosimetric and clinical data were normalized using log-transform before analysis. Kaplan-Meier curves were plotted by MedCalc and compared by log-rank test. R packages were used as follows: "glment" for LASSO-Cox regression, "Rms" for nomogram and calibration curves, "Hmisc" for correlation matrix, and "HemI" (Heatmap Illustrator, version 1.0) for heat map. "ipw" (Supplementary S2) was applied for estimating the inverse probability weights to control confounding (16).

Patient Characteristics
Patient characteristics and cycles of IC were compared between the CC and non-CC cohorts and between the AC and non-AC cohorts ( Table 1). No significant differences were observed between the two cohorts besides age (p = 0.001) and hemoglobin (p = 0.04) in the IC ± CC cohort. Forty (9.2%) patients were at stage II. After a median follow-up of 50 months (range, 2-113 months), 44 patients experienced locoregional recurrence, 64 patients experienced distant recurrence, and 99 patients died of NPC. For the whole dataset, the 5-year LRFS, DMFS, and OS rates were 85.5, 82.4, and 83.8%, respectively.

Association Between Dosimetric Parameters and Tumor-Related Data
In the IC ± CC cohort, the univariate Cox analysis identified T stage, N stage, TNM stage, and volume of GTVnd as tumorrelated risk factors of RFS ( Table 2). We then plotted the correlation matrix and heat map to illustrate the association between the nine dosimetric parameters and tumor-related data in Supplementary Figure S2 and Figure 2, respectively. The results showed that T stage was significantly associated with four dosimetric parameters, including PGTVnx_CI, PGTVnx_HI, PGTVnx_D98, and PGTVnx_D99 (all p ≤0.001). N stage and GTVnd_Volume were associated with the same five dosimetric parameters, including PGTVnd_0.3ccmin, GTVnd_4.8cc, PGTVnd_0.6ccmin, GTVnd_1.2cc, and GTVnd_4.7cc (all P ≤0.05). TNM stage was associated with six dosimetric parameters, including PGTVnx_D99, PGTVnx_D98,

Development of the Prognostic Nomogram
A subsequent multivariate Cox analysis showed that dosimetric signature (HR, 4.278, 95% CI: 2.36-7.755) and TNM stage (HR, 1.944, 95% CI: 1.161-3.255) were two independent risk factors for RFS ( Table 2). As age is a key consideration in the NPC treatment, a nomogram incorporating age, TNM stage, and dosimetric signature to predict the 3-and 5-year RFS was developed in the IC ± CC cohort ( Figure 3A).

Chemotherapy in Patient Groups With Low and High Risk
We performed a subset analysis to evaluate the survival benefit of CC in low-and high-risk groups. According to the optimal cutoff risk score of 0.60 identified by the X-tile plot, the patients were stratified into low-risk group (<0.60) and high-risk groups (≥0.60). The Kaplan-Meier curves showed significantly different LRFS, DMFS, and OS between the two groups (Supplementary Figure  S3). In the low-risk group, IC + CCRT was superior to IC + IMRT in terms of LRFS (HR, 0.27; 95% CI: 0.062-1.248; P = 0.03) and OS (HR, 0.363; 95% CI: 0.111-1.184; P = 0.036) but not DMFS (HR, 0.598; 95% CI: 0.216-1.651; P = 0.26) ( Table 3 and Figures 4A-C). CC was associated with a 9.4% increase of 5-year LRFS and an 8.8% increase of 5-year OS (Supplementary Figure S4). IC + CCRT + AC was not superior to IC + IMRT + AC in terms of LRFS, DMFS, and OS (P = 0.145, 0.506, and 0.093, respectively) ( Table 3 and Figures 4G-I). In the high-risk group, IC + CCRT was not superior to IC + IMRT in terms of LRFS, DMFS, and OS (P = 0.2, 0.874, and 0.668, respectively); IC + CCRT + AC was not superior to IC + IMRT + AC in terms of OS (HR, 4.735; 95% CI: 1.784-12.566; P = 0.022) and LRFS and DMFS (P = 0.148 and   We further explored the additional survival benefit of AC in both the low-and high-risk groups. In the low-risk group, IC + IMRT + AC was not superior to IC + IMRT in terms of   Table S1 and Supplementary Figurea S5D-F), AC was associated with a 33.6% increase of 5-year LRFS (Supplementary Figure S4), and IC + CCRT + AC was not superior to IC + CCRT in terms of LRFS, DMFS, and OS (P = 0.757, 0.779, and 0.964, respectively) (Supplementary Table S1 and Supplementary  Figures S5J-L).

DISCUSSION
Based on dosimetric parameters derived from the IMRT plan in patients with stage II-IVa NPC treated with IC followed by IMRT, we constructed a novel nomogram to quantify the risk of RFS of individuals. Using the optimal cutoff risk score calculated by the model, we successfully categorized the patients into low-and highrisk subgroups with significantly different survival outcomes. Patients with a low risk could benefit from CC by improving the 5-year LRFS and OS, while patients with a high risk could choose IC + IMRT + AC to obtain an improved 5-year LRFS. A total of 808 dosimetric factors were extracted and reduced to a panel of nine potential prognosticators by using a LASSO-Cox. LASSO is a robust regression technique for penalized analysis and variable selection from the mass data of dosimetric parameters (14,17,18). Cox regression analysis recognized the dosimetric signature, composed of nine factors, as an independent risk factor of RFS even after adjustment for age, TNM stage, and other confounders. Dosimetric signatures were essentially generated from radiotherapy plans based on contours that are a product of disease extent (TNM stage and/or disease volume). As potential confounders (19), disease extent should be incorporated into a regression analysis ( Table 2). Unexpectedly, it was TNM stage, not disease volume, that associated with RFS. The reasons for this result can be explained by the inadequate tumor coverage (reflected by the dosimetric signature) that were mainly from TNM stage (adherence to adjacent critical structures) rather than disease volume in the IMRT era (8). It is clinically challenging to strike a balance between the risk of recurrence due to inadequate tumor coverage and radiation complications of the adjacent critical structures.
Considering that inadequate tumor coverage is mainly from the disease extent in the IMRT era, the association between dosimetric signature and tumor characteristics was investigated in this study. A higher T stage was related to higher conformal index and homogeneity index. With the primary tumor infiltrating deeply, the margins between gross tumor and nearby critical OARs become narrower and narrower. Because of the radiation safety principle "as low as reasonably practicable" for critical OARs in clinical practice (8), a steeper dose fall-off outside the target volume and a better conformal index are prone to be observed in advanced T stage patients (20,21). As a compromise to the critical OARs, inhomogeneous target dose distribution is a strategy without an apparent dosimetric inadequacy of target volumes (22). However, inhomogeneous plans may be attributable to more cold spots (23), as the representative factors of cold spots, PGTVnx_D98 or PGTVnx_D99, seem to be ignored by PGTVnx_D95 in plan evaluation (24). As for N stage and GTVnd_Volume, we found that all the dose factors were associated with the two regardless of the volume and dose. It may be due to the progressive deformation of neck soft tissue structures from intrafraction motion (25). In addition, a recent study suggested that positive regional lymph nodes should be focused prior to primary tumor in NPC radiotherapy (14).
Our dosimetric signature does not contain D95 that is often used in the radiotherapy area. A plausible explanation for this might be that more attention on D95 in daily clinical practice limits its power. Definitive chemoradiotherapy is recommended by the 2021 National Comprehensive Cancer Network guidelines for stage II-IVa NPC patients. Because the quality of IMRT plans mainly relies on the experience of the dosimetrists (26), our dosimetric signature performed individualized risk quantification of IMRT plan for RFS. making it a surrogate for plan quality. It is therefore plausible that the chemotherapy benefit may be making up for poor radiation planning.
We found that CC was not a significant prognosticator of survival outcomes in the multivariate analysis. However, a subgroup analysis showed that CC can benefit patients with low risk instead of high risk in improving LRFS and OS. This finding is consistent with a gene expression-based signature study (27), a clinical experience-based report (28), and a doseeffect analysis (29). However, the results differed from other previous studies (30,31), possibly because of the patients treated with 2D (31) or no comparison of radiotherapy alone in the IMRT subgroup analysis (30). We expected the results of an ongoing phase 3 clinical trial (NCT02633202) comparing CCRT with IMRT alone to establish the role of CC in the IMRT era.
Our study showed that AC can benefit high-risk patients treated with IC + IMRT in terms of LRFS. The modality of IC + IMRT + AC may be a promising strategy for the high-risk group in the IMRT era. A phase III trial proved that AC with intravenous cisplatin and 5-fluorouracil failed to improve failurefree survival and OS after CCRT but increased the toxic effects in NPC (32). However, efforts have been made to find the candidates for AC. A postradiotherapy plasma Epstein-Barr virus DNA screening failed to identify patients with positive treatment effects (33). Interestingly, a benefit of distant failure and overall survival was obtained in high-risk patients with modified oral chemotherapy agents (34,35). Compared to platin-based chemotherapy, this kind of metronomic chemotherapy was associated with less toxicities, good compliance, and potential activation of immunity (36). A 2019 American Society of Clinical Oncology abstract (abstract no. 6046) reported that the total cumulative cisplatin dose (concurrent/induction/adjuvant) in multimodality therapy may be an independent prognosticator of LRFS and OS in locally advanced NPC. Therefore, the effect of CC was overemphasized, whereas the effect of AC was underestimated (34). Platin-based chemotherapy should be used post-IMRT instead of during IMRT in high-risk patients.
This study also has some limitations. Firstly, this was a retrospective study performed in a single center. Despite extensive adjustment, it is still possible that some amount of unmeasured confounders remains. Secondly, we did not consider the cumulative cisplatin dose because there was no consensus on the optimal cumulative cisplatin dose (29). Thirdly, EBV-DNA and its response after treatment is a major determinant of RFS and OS in patients with NPC (33,37). However, the EBV-DNA in our datasets was measured at different time-points (e.g., post-IC, post-IMRT). A future update of the nomogram including the EBV data would strengthen the findings. Fourthly, setup error was not taken into consideration as a reason of limited daily CBCT data; a future study is needed to investigate the dosage inconsistency between treatment plan dose and the actual delivered dose which resulted from the error based on CBCT data. Finally, the clinical use of the nomogram can be tested if an external validation is executed.

CONCLUSIONS
According to our knowledge, this is the first study to explore the role of CC in the viewpoint of IMRT-based dosimetric factors. Our findings revealed that the nomogram can effectively categorize patients with stage II-IVa NPC into subgroups with low and high risk of survival outcomes. A further analysis found that CC could improve LRR and OS for patients at low risk. Significantly, we identified that the high-risk patients could benefit from AC instead of CC. Prospective trials are needed to verify the positive treatment effect of CC and AC in low-r and high-risk patients, respectively. Our nomogram will allow clinicians to make more informed decisions for better patient care.

DATA AVAILABILITY STATEMENT
The original contributions presented in the study are included in the article/Supplementary Material. Further inquiries can be directed to the corresponding authors.

ETHICS STATEMENT
The studies involving human participants were reviewed and approved by the Ethics Committee on Biomedical Research, West China Hospital of Sichuan University. The ethics committee waived the requirement of written informed consent for participation.

ACKNOWLEDGMENTS
We acknowledged all faculty members, staff, and students of our department. We also gratefully acknowledged Ms. Yunqi Zhong for her valuable suggestions and discussions.

SUPPLEMENTARY MATERIAL
The