Prognostic nomogram for refining prognostication of the proposed AJCC/UICC 8^th edition for nasopharyngeal cancer in the era of intensity-modulated radiotherapy

Abstract

Objective

To develop a nomogram for refining prognostication for patients with non-disseminated nasopharyngeal cancer (NPC) staged with the proposed AJCC/UICC 8^th edition.

Material and methods

Consecutive patients investigated by magnetic resonance imaging, staged by the proposed AJCC/UICC 8^th edition, and irradiated by intensity- modulated radiotherapy (IMRT) from June 2005 to December 2010 were analyzed. The cohort of 1197 patients treated at Fujian Provincial Cancer Hospital was used as the training set and the results were validated by 412 patients from Pamela Youde Nethersole Eastern Hospital. Cox regression analyses were performed to identify significant prognostic factors for developing a nomogram to predict overall survival (OS). The discriminative ability was assessed with concordance index (C-index). Patients were categorized into three risk groups by performing recursive partitioning algorithm (RPA) on the survival scores of the combined set.

Results

Multivariable analysis showed that age, gross primary tumor volume (GTV-P) and lactate dehydrogenase (LDH) were independent prognostic factors for OS in addition to stage-group. The OS nomogram based on all these factors had a statistically higher bias-corrected C-index than prognostication based on stage-group alone (0.712 vs 0.622, p<0.01). These results were consistent for both the training and the validation cohorts. Patients with <135 points were categorized as low-risk, ≥135–<160 points as intermediate-risk and ≥160 points as high-risk, respectively. Their 5-year OS rates were 92%, 84% and 58%, respectively.

Conclusions

The proposed nomogram could improve prognostication when compared with TNM stage-group. This could aid in risk stratification for individual NPC patients.

Introduction

Nasopharyngeal carcinoma (NPC) is a peculiar cancer that is endemic in Southeast Asia. The natural behavior and therapeutic considerations for NPC are different from other head and neck cancers. Radiotherapy (RT) is the primary treatment modality; chemotherapy is added to radiation in patients with locoregionally advanced disease. For this highly infiltrative cancer anatomically surrounded by critical structures, conformal RT techniques are needed for optimal coverage of tumor extent with better protection of normal tissues. The technology of RT has evolved substantially during the past decades; intensity-modulated technique (IMRT) is the current standard of care if resources are available.

Accurate prediction of failure pattern and survival outcome is crucial for patient management. TNM stage is presently the most important known prognostic factor. Besides meeting the demand for accurate prognostication, the ideal staging system must also be globally applicable. The American Joint Committee for Cancer Staging (AJCC) and the International Union against Cancer (UICC) have been taking concerted efforts to continually improve the TNM Staging system in line with evolving investigation and treatment methods. The AJCC/UICC 8^th edition will soon be introduced [1]; multidisciplinary international experts are involved in the preparatory process. The proposed system is based not only on extensive literature review, but also on validation of recommendations by a contemporary series of 1609 patients who were staged with MRI and irradiated with IMRT from 2 hospitals (one in Hong Kong and the other in Fujian in Mainland China).

It is well recognized that the TNM system does have limitations because it is entirely based on anatomical disease extension [2]. Variability in outcome among patients within the same stage receiving similar treatments is commonly observed; this could be attributed to heterogeneity in patient factors and other tumor factors. Literature review suggests the following potential alternative or additional prognostic factors may be useful: age [3, 4], gender [3], performance status [5], gross primary tumor volume (GTV-P) [6–9], lactate dehydrogenase (LDH) level [10–12], deoxyribonucleic acid (DNA) copies of Epstein-Barr Virus (EBV) [13–15], hemoglobin level [16]. Among these factors, EBV-DNA is a promising parameter, but not routinely available for all patients in most centers; the test methodology is difficult and requires further harmonization for global application [17]. The aim of the current study is to identify commonly available factors with independent statistical significance and to explore development of a prognostic nomogram by combining TNM parameters with these factors for further refinement of prognostication for individual NPC patients.

Methods

Patients Characteristics

Consecutive patients with non-disseminated NPC treated with curative intent from Fujian Provincial Cancer Hospital (mainland China) and Pamela Youde Nethersole Eastern Hospital (Hong Kong, China) from June 2005 to December 2010 were analyzed (Table 1). All patients had histological confirmation. They were staged with MRI and irradiated with IMRT. None had history of previous treatment or prior malignancy. The series of 1197 patients from Fujian Provincial Cancer Hospital was used as the training cohort. Their median age was 46, with a male to female ratio of 3:1. The majority of patients had advanced locoregional diseases: 70% were categorized as stage III-IVA by the proposed AJCC/UICC 8th edition [1]. The corresponding series of 412 patients from Pamela Youde Nethersole Eastern Hospital was used as the validation cohort. Table 1 shows the characteristics of the two cohorts of patients.

Table 1.

Patient characteristics

Number (%)	Training Cohort (N=1197)	Validation Cohort (N=412)	P value
Age (years)			<0.001
Median (range)	46 (11–84)	50 (17–84)
Mean	46.4	51.1
Gender			0.706
Female	292 (24.4)	105 (25.5)
Male	905 (75.6)	307 (74.5)
Performance status			<0.001
0	1087 (90.8)	337 (81.8)
1	107 (8.9)	65 (15.8)
2–3	3 (0.3)	10 (2.4)
T-category*			<0.001
1	285 (23.8)	57 (13.8)
2	220 (18.4)	66 (16.0)
3	294 (24.6)	191 (46.4)
4	398 (33.2)	98 (23.8)
N- category*			<0.001
0	174 (14.5)	20 (4.9)
1	658 (55.0)	53 (12.9)
2	270 (22.6)	209 (50.7)
3	95 (7.9)	130 (31.6)
Group stage*			<0.001
I	57 (4.8)	6 (1.5)
II	297 (24.8)	26 (6.3)
III	381 (31.8)	184 (44.7)
IVA	462 (38.6)	196 (47.6)
Histology			0.328
WHO Type 1	12 (1.0)	1 (0.2)
WHO Type 2.1	51 (4.3)	17 (4.1%)
WHO Type 2.2	1134 (94.7)	394 (95.6)
GTV-P (cc)			0.053
Median (range)	32.8 (0.1–235.6)	31.1 (0.7–279.9)
Mean	41.2	46.1
LDH (IU/L)			<0.001
Median (range)	183 (106–751)	140 (71–859)
Mean	193.4	148.1
Hemoglobin (g/dL)			0.006
Median (range)	14.3 (8.0–17.1)	14.4 (7.1–17.9)
Mean	14.3	14.0
Radiotherapy dose (Gy)			0.960
Median (range)	69.75 (61.60–86.65)	70 (66–70)
Mean	69.98	69.98
Addition of chemotherapy
Overall (any sequence)	1016 (84.6)	343 (83.3)	0.479
Concurrent ± Non-concurrent	467 (39.0)	338 (82.0)	<0.001

Performance status: based on ECOG system;

^*Proposed 8th Edition of AJCC/UICC Staging System; GTV-P = gross primary tumor volume; LDH – lactate dehydrogenase; WHO Type 1 = Keratinizing, WHO Type 2.1 = Non-Keratinizing (differentiated), WHO Type 2.2 = Non-Keratinizing (undifferentiated)

Clinical Staging and Treatment

All patients had complete physical examination, fiberoptic nasopharyngoscopy and MRI of the nasopharyngeal and cervical regions. Additional metastatic evaluation was performed in accordance with institutional polices. All of the patients were re-staged using the criteria of the proposed AJCC/UICC 8^th edition [1]. All patients were irradiated with IMRT technique. The GTV-P was first contoured by a diagnostic radiologist on contrast enhanced MRI. After rigid fusion with contrast enhanced planning computer tomography, the GTV-P was modified and delineated by a radiation oncologist according to the fused images. The median total dose was 69.8Gy (range, 61.6–86.7). Details of IMRT planning and dose prescription have been described previously [18, 19]. Additional treatment with cisplatin-based chemotherapy (various schedules) was administered to 85% of the training cohort and 83% of the validation cohort (Table 1).

Statistical analysis

All events were measured from the date of histological diagnosis. The primary end-point in the current study was overall survival (OS) which was defined as the interval from histological diagnosis to date of death due to any cause. Survival rates were calculated by the Kaplan-Meier method and were compared using log-rank test. Two-sided tests were used and those with p value <0.05 were considered statistically significant. Potential factors routinely evaluated in the current cohorts (including age, gender, performance status, GTV-P, LDH level, hemoglobin level) were analyzed by the Cox’s proportional hazards model. The hazard ratio (HR) with 95% confidence level (CI) was assessed. Variables were tested in both univariate and multivariate analyses. The final model was obtained by a backward selection; and this was used to generate the nomogram for predicting 5-year and 8-year survival rates [20]. The discrimination ability was evaluated by the concordance index (C-index) with bias-correction. Internal validation was performed by bootstrap algorithm and 1000 bootstrapping replications were generated to provide the bias-corrected C-index and corresponding 95% credible intervals. The calibration plot comparing the nomogram predicted versus observed Kaplan-Meier estimates of the survival probability was used to assess the accuracy. To test for generalizability, the proposed nomogram developed from the training cohort was tested with the validation cohort.

Patients were categorized into three risk groups (low, moderate and high risk) by performing recursive partitioning algorithm (RPA) [21] on the survival probability scores of the whole series. Statistical analyses were performed using software R version 3.13, with packages rms and rpart for constructing the nomogram and performing the RPA, respectively.

Results

Among the training cohort, 219 (18%) patients had died, the 5-year OS rate was 82 %. Univariate analysis on OS showed that age, gender, performance status, stage-group, GTV-P and pre-treatment LDH level were all associated with survival outcome (p<0.05). Hemoglobin level was not statistically significant. Multivariate analyses showed that in addition to stage-group, three more factors also showed independent significance: age, GTV-P and LDH level (Table 2).

Table 2.

Significant factors on overall survival by multivariate analysis of the training cohort (N = 1197)

Variables	HR (95% CI)	p
Stage-group		0.002
II vs I	2.58 (0.61, 10.83)	0.196
III vs I	4.67 (1.15, 18.98)	0.031
IVA vs I	5.86 (1.43, 24.06)	0.014
Age (per year increase)	1.04 (1.03, 1.06)	<0.001
GTV-P (per cc increase)	1.008 (1.004, 1.013)	<0.001
LDH (per IU/L increase)	1.005 (1.002, 1.008)	0.002

Gender and performance status were not statistically significant in the model.

Development of nomogram

A nomogram predicting the 5-year and 8-year OS rates was generated using these four prognostic factors with their relative contributions (Figure 1). Different weighted points were assigned to the different factors as continuous variables. Weighted total point was calculated by summing up the point of each variable, which was then vertically aligned to the 5-year and 8-year OS rates. Internal validation of the proposed nomogram using bootstrapping resampling revealed a bias-corrected C-index of 0.712 [median, 95% credible interval (CrI): 0.674–0.747]. This was statistically higher than that of prognostication based on stage-group alone (0.622, 95% CrI: 0.590–0.653, p <0.01), indicating superior discrimination ability. The calibration plot of the proposed nomogram also showed a perfect agreement between the predicted versus observed Kaplan-Meier estimates of the OS probability (Figure 2A, intercept=−0.05, p=0.60; slope=1.06, p=0.59).

Figure 1.

Nomogram for 5-year and 8-year overall survival (OS) based on the training cohort

Figure 2.

Calibration plot for overall survival on (A) training cohort (intercept=−0.05, p=0.60; slope=1.06, p=0.59) and (B) validation cohort (intercept=0.15, p=0.25; slope=0.89, p=0.47)

Repeating the analyses using the validation cohort showed consistent improvement with the proposed nomogram: the bias-corrected C-index was 0.760 (95% CrI: 0.723–0.796); the superiority in discriminating ability as compared with stage-group alone [bias-corrected C-index 0.654 (95% CrI: 0.622–0.686)] was confirmed (p<0.01). There was good agreement between the predicted and observed OS estimates (Figure 2B, intercept=0.15, p=0.25; slope=0.89, p=0.47).

Risk-Groups Categorization

Basing on the scores estimated from the developed nomogram on OS, the whole series of patients could be categorized into three risk groups by RPA (Figure 3). Rounding the classified cutoff scores, 45% patients with <135 points were categorized as low-risk, 27% patients with ≥135–<160 points as intermediate-risk, and 29% patients with ≥160 points as high-risk. Their 5-year OS rates were 92%, 84% and 58%, respectively. The c-index was 0.705 (95% CrI: 0.676–0.730).

Figure 3.

Recursive partitioning algorithm for overall survival of the whole series based on the score derived from the nomogram

Basing on the proposed 8^th edition of the AJCC/UICC staging system [1], the distribution pattern was: 4% stage I, 20% stage II, 35% stage III, 41% stage IVA. Their 5-year OS rates were 98%, 92%, 83% and 71%, respectively. The c-index was 0.628 (95% CrI: 0.602–0.653). Comparison of the proposed nomogram versus the 8^th edition AJCC/UICC staging system regarding distribution pattern, OS at 5-year and the c-index of risk-groups were summarized in Table 3, and the respective Kaplan-Meier estimates on OS were shown on Figure 4. The comparison of c-index suggested that the risk groups based on nomogram (taking into account all significant prognostic factors) had significantly better differentiating ability than stage-group alone (p<0.01).

Table 3.

Comparison of risk-group based on nomogram versus stage-group per se for the whole series (N = 1609)

	Stage-group by AJCC/UICC system*		Risk-group by nomogram
Distribution	I:	3.9%	Low:	44.6%
	II:	20.1%	Intermediate:	26.9%
	III:	35.1%	High:	28.5%
	IVA:	40.9%
Overall survival (5-year)	I:	98.4%	Low:	91.8%
	II:	91.5%	Intermediate:	84.2%
	III:	83.0%	High:	58.4%
	IVA:	71.1%
C-index, median (95% CrI)	0.628 (0.602–0.653)		0.704 (0.676–0.730)

^*(Proposed 8^th edition);

Risk-groups based on nomogram: low-risk, 0–135; intermediate-risk, >135–<160; high-risk, ≥160 points.

Figure 4.

Kaplan-Meier plots for overall survival of the whole series (A) Stage-group by AJCC/UICC Proposed 8^th Edition, (B) risk-groups by nomogram

Discussion

Nomograms are pictorial representations of complex mathematical relationships. They enable incorporation of both patient and disease characteristics into the calculation of a simple numerical estimate of the event probability. In addition to the anatomical disease extent, nomograms incorporate other key prognostic factors into the estimation of survival outcomes. These can tremendously facilitate individualized risk stratification and decision making [22–24]. The utilization of nomograms has been well recognized in other cancers and has been shown to yield more accurate survival predictions over the traditional TNM staging system [25–30]. These findings underlined the inadequacy of the current staging system and the unmet need of a more robust prognostic tool for risk categorization of individual patients.

Six studies on development of nomograms for NPC have been reported in recent years [5, 31–35]. They demonstrated feasibility of using nomogram to predict survival in NPC and consistently achieved superiority to the conventional TNM staging system. However, despite the stronger predictive power of these nomograms, they suffered from important limitations including obsolete RT techniques [33–35] and lack of external validation [31, 33] that hampered their generalizability and validity in modern era. Furthermore, all were based on the 6–7th editions of AJCC/UICC staging system [5, 31–35]. With the imminent introduction of the 8^th edition, applicability of these nomograms will become outdated.

Five of these nomograms were developed using patient data from the same cancer institute in China [31–35]. The nomogram by Tang et al. [33] was based on a large cohort of 4630 patients, but only 34% of patients were treated with IMRT. In addition to common prognostic factors (age, gender, T-category, N-category, LDH level and hemoglobin), factors not observed in other centers [body mass index (BMI), C-reactive protein (CRP)], and factors not routinely available in majority of centers (plasma EBV-DNA) were incorporated into the proposed nomogram. Furthermore, all the continuous variables were transformed into categorical data in the statistical analyses. Although all these prognostic factors were statistically significant in multivariate analyses, the uncertainties of the best cut-off value and the issue of over-fitting in the model should be carefully considered.

The nomogram proposed in the current paper is the first based on the 8^th edition of the AJCC/UICC TNM staging system [1] that soon will be implemented. The whole series of 1609 patients were all staged and treated with current standard of care (using MRI and IMRT). The prognostic factors included are well recognized and routinely captured by contemporary centers (independent factors identified being age, GTV_P and LDH level). Moreover, two data cohorts during the same study period were analyzed so that the nomogram based on the primary institute was validated by data from another institute to ensure its robustness and generalizability.

Among the host factors, age was a powerful prognostic factor for OS [36–38]. Elderly patients had poorer survival partly due to comorbidities and partly due to poorer performance status leading to a lower compliance to intensive treatment [4]. The current study did not show substantial significance attributed to gender or performance status.

Among the disease factors, there are increasing data suggesting that tumor volume and LDH have independent significance that could supplement prognostication by the current TNM staging system that only takes into account the anatomical extent of gross disease. A larger tumor burden often signifies an increase in the number of clonogenic tumor cells and radio-resistance associated with tumor hypoxia. Sze et al. [6] estimated that the local failure would increase by 1% for every 1 ml increase in tumor volume. Wu et al. [7] reported 48 ml and 25 ml were the GTV-P cut-off points for local control and distant metastasis respectively. Willner et al. [8] found that tumors larger than 64 ml had poorer control rate. Among all these studies, the series by Guo et al. [9] was the largest (n=694) and their results suggested that incorporation of tumor volume could improve the prognostic ability of T-category. Our finding concurred with these publications. In the past, there are concerns that measurement of GTV-P was not easy or routinely available. However, with the contouring of gross tumor volume for all patients irradiated with modern RT technique, this factor is now captured by computerized planning system for all patients (except those treated with 2D technique). Hence, refining prognostication by incorporating GTV-P into the nomogram has become widely achievable.

The current study concurred with other reports that high pre-treatment LDH level was an adverse prognostic factor, associated with increased distant metastasis [10–12, 31, 33–35]. A retrospective study done by Wan et al. (32) showed that patients with high pre-treatment LDH > 177.0 U/L had significantly inferior 5-year distant metastasis free survival as compared with those with ≤ 177.0 U/L (65% vs 77%, p =0.016). As LDH is easily measurable in all centers, this should be incorporated for refining prognostication.

The currently proposed nomogram based on incorporation of age, GTV_P and LDH in addition to stage group by the new TNM 8^th edition [1], showed significantly better discrimination ability as compared with stage group alone (p <0.01): the median bias-corrected C-index was 0.712 (95% CrI: 0.674–0.747) by nomogram versus 0.622 (95% CrI: 0.590–0.653) by stage.

Besides risk estimation for individual patient by nomogram, we further explore the feasibility of using the proposed nomogram to categorize patients into more accurate risk groups. By performing RPA based on the scores estimated from the proposed nomogram on OS, 45% patients with <135 points were categorized as low-risk, 27% patients with ≥135–<160 points as intermediate-risk, and 29% patients with ≥160 points as high-risk. Their corresponding 5-year OS rates were 92%, 84% and 58%, respectively. The c-index was 0.705 (95% CrI: 0.676–0.730). Not only does this risk-group categorization exhibit superior differentiating ability than the TNM stage-group alone, but it also achieves a more even pattern of patient distribution, as compared with stage-group distribution of 4% stage I, 20% stage II, 35% stage III, 41% stage IVA.

A more precise risk stratification system is always desirable in patient counseling and treatment decision. Furthermore, this is valuable for future clinical trial design. Further studies based on this risk categorization to evaluate the possibility of treatment de-escalation for the low-risk group and treatment intensification for the high-risk group will be valuable. This refined prognostic tool is fundamental for working towards personalized treatment strategies according to individual risk.

Nevertheless, our study did have several limitations. Firstly, it is a retrospective study, predominantly on patients with non-keratinizing subtype of NPC. Prospective validation in non-endemic centers will be desirable. Secondly, the measurement of GTV-P depends heavily on the quality of diagnostic imaging and the current contouring could be operator dependent. Further enhancement of accuracy in delineating GTV-P by standardized imaging protocols and computerized program are be desirable.

Another potentially valuable prognostic factor to be considered is plasma EBV DNA. There are increasing data suggesting that it is an important marker with strong correlation with tumor burden, TNM stage, response to treatment and survival [13–15]. However, this is not yet routinely available in most centers worldwide. The measurement of EBV DNA is not easy; harmonization of the methodology for quantitative plasma EBV DNA assay is currently being explored by international collaboration [17] It is also an expensive test and the optimal cut-off value has yet to be defined.

In conclusion, the currently proposed nomogram for NPC patients provides a valuable tool for refining prognostication and working towards personalized medicine. This is the first proposed system for improving the differentiating ability and distribution pattern of stage group by the AJCC/UICC TNM 8^th edition [1] that will soon be introduced. This is applicable to patients staged and treated with modern standard of care; the incorporated factors are consistently recognized as significant factors and globally available (except in centers without 3D/IMRT facilities for measuring GTV-P). Further confirmation through larger scale prospective multi-institutional collaborative studies is warranted. Continuous efforts for further improvement by incorporation of novel markers will be valuable. Lastly, development of an easily operated smartphone app can be considered to enhance the usage of the prognostic model.

Condensed abstract: A nomogram incorporating TNM stage-group, age, GTV-P volume and LDH significantly improves the prognostic power when compared with TNM stage-group alone. More personalized treatment decisions can be achieved with this refinement of risk stratification for individual NPC patients.

Condensed abstract.

A nomogram incorporating TNM stage-group, age, GTV-P volume and LDH significantly improves the prognostic power when compared with TNM stage-group alone. More personalized treatment decisions can be achieved with this refinement of risk stratification for individual NPC patients.