Nomograms for preoperative prediction of prognosis in patients with oral cavity squamous cell carcinoma


  • Presented at the 8th International Meeting of the American Head and Neck Society in Toronto, Canada, July 21-25, 2012.



This study sought to develop prognostic tools that will accurately predict overall and cancer-related mortality and risk of recurrence in individual patients with oral cancer based on host and tumor characteristics. These tools would take into account numerous prognosticators beyond those covered by the traditional TNM (tumor–node–metastasis) staging system.


Demographic, host, and tumor characteristics of 1617 patients with cancer of the oral cavity, who were treated primarily with surgery at a single-institution tertiary care cancer center between 1985 and 2009, were reviewed from a preexisting database. Recurrent disease was recorded in 509 patients (456 locoregional and 116 distant); 328 patients died of cancer-related causes, and 542 died of other causes. The median follow-up was 42 months (range, 1-300 months). The following variables were analyzed as predictors of prognosis: age, sex, race, alcohol and tobacco use, oral cavity subsite, invasion of other structures, comorbidity, tumor size, and clinical nodal status. The stepdown method was used to select the statistically most influential predictors for inclusion in the final nomogram for each outcome of interest.


The most influential predictors of both recurrence and cancer-specific mortality probability (CSMP) were tumor size, nodal status, subsite, and bone invasion. Nomograms were generated for prediction of overall survival (OS), CSMP, and locoregional recurrence-free probability (LRRFP). The nomograms were internally validated with an overfit-corrected predictive discrimination metric (concordance index) for OS of 67%, CSMP of 66%, and LRRFP of 60%.


Nomograms have been developed that can reasonably estimate OS, CSMP, and LRRFP based on specific tumor and host characteristics in patients with oral cancer. Cancer 2014;120:214–221. © 2013 American Cancer Society.


The TNM (tumor–node–metastasis) staging system is an effective and user-friendly tool for prognostic prediction in patients with oral cavity cancer. It offers a simple method to estimate the prognosis of patients with cancer based on specific anatomic characteristics of the tumor. The TNM system has been widely used to plan treatment, summarize prognostic information, evaluate treatment results, and compare outcomes between institutions around the world.[1] However, the main drawback of the TNM system has been its inability to adapt to advances in our understanding of cancer biology and incorporate new prognostic variables as they become available.[1] It does not account for the prognostic influence of other variables such as the anatomic subsite and many host characteristics such as the presence of comorbidities and lifestyle behaviors. This restriction originates primarily from the rigid “bin” configuration of TNM, which means that any attempt to include new variables or categories exponentially increases the number of bins, multiplies the stage options, and makes the system unwieldy.[2] Furthermore, the outcome of interest used by the TNM staging system is overall survival (OS); however, other outcomes, such as disease-specific survival, local control, or regional control, may be more relevant to assess treatment results and different therapeutic options.[3]

A statistical tool such as a nomogram has the ability to take into account numerous variables to predict an outcome of interest for an individual patient.[4] Nomograms have been widely tested in a variety of cancers, with more than 1000 publications available in the literature. In contrast to these other cancers, nomograms have been used only sparingly for head and neck tumors. Nomograms will likely find widespread use in cancer prognostication, especially with the availability of good quality data and increasing sophistication in computing software and technology. We have previously outlined a modular concept to use nomograms in the staging of head and neck cancer.[3] As an initial step toward achieving this goal, we report the development of a set of nomograms that can accurately predict an individual patient's OS, risk of cancer-related mortality, and risk of locoregional recurrence, using information available prior to surgical management, while maintaining the user friendliness of the TNM system.


After Institutional Review Board approval was obtained, a retrospective cohort study of patients treated for squamous cell carcinoma of the oral cavity (SCCOC) at Memorial Sloan-Kettering Cancer Center (MSKCC) was conducted. Patients were excluded from the study if they had previous head and neck cancer, radiotherapy, chemotherapy, M1 disease, or primary tumors of the lip. Between January 1985 and December 2009, 1617 patients were included in this study. The last follow-up by a head and neck surgeon at MSKCC was recorded in September 2011.

Patients and tumor characteristics were recorded from clinical charts: age, sex, race, alcohol use, tobacco use, presence of comorbidities, subsite of the tumor within the oral cavity, largest dimension of the primary tumor on clinical examination/radiographic imaging, presence of bone, skin, or deep tongue muscle infiltration, and the clinical/radiographic status of the neck lymph nodes. Alcohol and tobacco use was categorized as current versus never/former. Medical comorbidities were classified using the Washington University Head and Neck Comorbidity Index (WUHNCI)[5]; this scale assigns a score based on the presence of previously defined severe comorbidities (score ≥ 1). The subsites of oral cavity were those used by the American Joint Committee on Cancer (AJCC) TNM staging system.[6] The largest dimension of the primary tumor and presence of bone, skin, and deep tongue muscle infiltration were recorded at clinical examination and/or from imaging studies. Age, comorbidity index score, and largest dimension of the primary tumor were analyzed as continuous variables. The rest of the covariates were analyzed as categorical variables.


The main outcomes of interest were OS, cancer-specific mortality probability (CSMP), and locoregional recurrence-free probability (LRRFP). Our data were sufficiently mature to make predictions at the 5-year time point.

The event when calculating OS was death from any cause. Time to event was calculated in months from the date of surgery to the date of death. Patients who did not die were censored at the date of last follow-up. Thus, our OS prediction represents an individual patient's probability of survival 5 years from surgery.

The event when calculating CSMP was death with active SCCOC. Deaths from other causes were considered as competing risks of CSMP, because they impeded us from observing the occurrence of CSMP. Patients who did not die were censored at the date of last follow-up. Time to event was calculated in months from date of surgery to the date of death or the date of last follow up. CSMP represents an individual patient's probability of dying with active SCCOC within 5 years of surgery.

The event when calculating LRRFP was recurrence in the tumor bed and/or neck. Time to event was calculated in months from date of surgery to the date of locoregional recurrence. Patients who did not recur in the tumor bed or neck were censored at the date of last follow-up, regardless of distant recurrence or death. LRRFP represents an individual patient's probability of being free from locoregional recurrence 5 years from surgery.

Nomogram Development

Multivariable Cox proportional hazards regression models were used to model survival data and recurrence, and a multivariable competing risk regression was used for disease-specific death. Death from other causes was treated as a competing risk for cancer-specific mortality. Missing values in the predictors were multiply imputed using the multivariate imputations by chained equations (MICE) before conducting multivariable regression statistical analysis.[7] The following variables were analyzed as predictors of prognosis: age, sex, race, alcohol and tobacco use, oral cavity subsite, comorbidities, tumor size, invasion of other structures (skin, bone, or deep muscle of tongue), and clinical nodal status. The stepdown model reduction method[8] was used to select the statistically most influential predictors for inclusion in the final nomogram for each outcome of interest; predictors that did not help the overall prediction were eliminated. Restricted cubic splines were applied to continuous variables to allow nonlinear relationships as appropriate.

Preoperative nomograms were generated from the final selected models for prediction of OS, CSMP, and LRRFP. The nomograms were internally validated using 1000 bootstrap resamples to assess their predictive accuracies if these nomograms are applied to future users. The predictive accuracy was quantified by calculation of the concordance index (C-index) for each outcome of interest,[8] which is equivalent to the area under the receiver operating characteristic curve for the binary outcomes, but adapted for censored or competing risks data. It has a scale of 0.5 to 1 with 1 representing perfect discrimination and 0 for no discrimination ability. In addition, the calibration between the predicted and observed probabilities was visually checked. A P value < 0.05 was used to determine the statistical significance. All statistical analyses and graphics were carried out using the open source statistical software R-2.13.2 (R Development Core Team, 2012) with MICE, Design, and “cmprsk” packages added.


Patient demographic and tumor characteristics are listed in Table 1. Seventy-four percent of patients had a primary tumor smaller than 4 cm in largest dimension (cT1 and cT2), and 70% of patients had clinically N0 necks. Adjuvant postoperative radiotherapy was used in 32% of patients after surgical resection, and chemoradiation in 3% of patients.

Table 1. Demographics and Clinical Tumor Characteristics
Age, y, mean (standard deviation)61.36(14.57)
Age <60 y69443%
Age >60 y92357%
Not reported30%
No severe120975%
Alcohol use  
Tobacco use  
Year of diagnosis  
Buccal mucosa946%
Oral tongue79649%
Floor of mouth25916%
Hard palate423%
Upper gum1006%
Lower gum22014%
Retromolar trigone1066%
Clinical tumor stage  
Clinical neck stage  
cN 0112870%
cN +48930%

The median follow-up interval following surgical treatment was 42 months (range, 1-300 months). A total of 509 patients developed recurrent disease (206 local, 141 regional, 109 locoregional), and 117 patients developed distant disease, 53 of whom had distant metastasis without local or regional recurrence. A total of 328 patients died of cancer-related causes, whereas 542 died of other causes.

The 5-year OS of patients in this series was 61%. Using the stepdown model reduction, the strongest predictors of OS were age, race, presence of severe comorbidities (WUHNCI ≥ 1), use of tobacco, and clinical nodal status. Table 2 depicts the impact of various factors on OS in multivariable Cox proportional hazards analysis: age (hazard ratio [HR] = 1.8, P < .001), race (HR = 1.4, P < .001), tobacco use (HR = 1.3, P < .001), comorbidities (HR = 1.2, P < .001), primary tumor dimension (HR = 1.4, P < .001), and the presence of clinically positive lymph nodes (HR = 1.5, P < .001).

Table 2. Hazard Ratios From the Final Cox Proportional Hazards Regression Model
VariableOverall Survival5-Year CSMP5-Year LRRFP
HR (95% CI)PHR (95% CI)PHR (95% CI)P
  1. Abbreviations: CI, confidence interval; CSMP, cancer-specific mortality probability; HR, hazard ratio; LRRFP, locoregional recurrence-free probability; NS, nonsignificant.

  2. a

    Statistically significant with P < .05

  3. b

    Predictors nonselected with the stepdown method.

  4. c

    Hazard ratios of continuous variables were based on the contrast of the third quartile (Q3) versus the first quartile (Q1).

Age, y (72 : 52)c1.8 (1.6–2.0)<.001ab b 
SexbNSb b 
Raceb b b 
Asian : White0.8 (0.6–1.1).015a    
Black : White1.4 (1.1–1.9)     
Comorbidity scale (1 : 0) c1.2 (1.1–1.3)<.001ab b 
Alcoholb b b 
Tobaccob b b 
Current : none/former1.3 (1.2–1.5)<.001a    
Subsite   .001a <.001a
Buccal mucosa : tongueb 0.7 (0.4–1.2) 1.1 (0.8–1.6) 
Floor of mouth : tongueb 0.5 (0.4–0.8) 0.6 (0.5–0.8) 
Hard palate : tongueb 1.3 (0.7–2.3) 1.8 (1.2–2.8) 
Lower gum : tongueb 0.5 (0.4–0.8) 0.8 (0.6–1.0) 
Retromolar trigone : tongueb 0.7 (0.5–1.1) 0.7 (0.5–1.1) 
Upper gum : tongueb 0.9 (0.6–1.4) 1.2 (0.9–1.7) 
Primary size dimension (3.5 : 1.8)c1.4 (1.2–1.6)<.001a1.6 (1.3–1.9)<.001a1.2 (1.0–1.4).044a
Skin invasionb b b 
Bone invasionb 2.0 (1.5–2.8)<.001a1.7 (1.3–2.2)<.001a
Deep muscle invasionb b b 
Clinical nodal status1.5 (1.3–1.8)<.001a1.7 (1.3–2.1)<.001a1.5 (1.2–1.8)<.001a

The 5-year CSMP was 24%. The main factors significantly related to CSMP in the stepdown analysis were primary tumor dimension, subsite of the tumor, clinical nodal status, and bone invasion (Table 2).

The 5-year LRRFP was 62%. The stepdown analysis demonstrated that the factors predictive of worse CSMP were also predictors of worse LRRFP. Table 2 lists the factors analyzed for impact on LRRFP in multivariable analysis. Primary tumor dimension (HR = 1.2, P < .001), clinical nodal status (HR = 1.5, P < .001), the subsite of the tumor, and bone invasion (HR = 1.7, P < .001) were each predictors of worse LRRFP.

Preoperative nomograms to predict 5-year OS, CSMP, and LRRFP were developed using the results of the multivariable Cox proportional hazards or competing risks regression analysis for each outcome. The nomogram to predict 5-year OS was developed using the most influential predictors: age, race, tobacco status, comorbidities, tumor diameter, and clinical N status (Fig. 1). A weighted total score was calculated from these factors; this score was used to provide estimates of 5-year OS. The resulting model was internally validated using an overfit-corrected discrimination measure. The model demonstrated good discrimination for prediction of OS with a C-index for the OS nomogram of 0.67.

Figure 1.

Overall survival nomogram is shown. A nomogram is a user-friendly graphic representation of a complex multivariate model. It puts the predictive contribution of each covariate on a scale of points from 1 to 100. The strongest variable has the largest effect and it is assigned 100 points. The rest of the variables (continuous and categorical) are assigned a smaller number of points proportional to their size effect. The total point axis represents the maximum potential score, and it is associated with a probability. The total points accumulated by the covariants correspond to the predicted probability of an outcome of interest for an individual patient. Abbreviation: WUHNCI. Washington University Head and Neck Comorbidity Index.

Tumor size, nodal status, subsite, and bone invasion were the variables with highest predictive value for CSMP. The CSMP nomogram constructed using these parameters is presented in Figure 2. The C-index of the nomogram was 0.66. The nomogram for predicting 5-year risk of LRRFP was constructed with the same set of predictors as the CSMP nomogram presented in Figure 3. The concordance index of this nomogram was 0.60. Calibrations for all 3 outcomes appear to be satisfactory (Figs. 4, 5, 6).

Figure 2.

Cancer-specific mortality probability nomogram is shown.

Figure 3.

Locoregional recurrence-free probability nomogram is shown.

Figure 4.

Overall survival nomogram calibration plot is shown. Accuracy of the nomogram is measured through the calibration plot and concordance index. This plot is a graphic representation of the nomogram calibration ability. The 45-degree line represents ideal predictions. The calibration curve represents how far the nomogram predicted probabilities are from the actual outcome in absolute risk scale. A nomogram predictive accuracy is measured via a concordance index (c-index), which quantifies the level of concordance between predicted probabilities and the actual chances of having the event of interest for patients with versus without outcomes.

Figure 5.

Cancer-specific mortality probability nomogram calibration plot is shown.

Figure 6.

Locoregional recurrence-free probability nomogram calibration plot is shown.


The utility of nomograms in individualized risk prediction relative to the current gold-standard TNM staging-based prognostic model is well recognized in a variety of cancers, such as sarcomas and breast and prostate cancer.[9-11] We have previously reported on the use of a nomogram for postoperative management of patients with oral cancer,[12] and proposed the use of nomograms for staging of head and neck cancer.[3]

With this study, we introduce the concept of using individualized risk prediction tools for estimating OS, cancer-specific mortality, and locoregional recurrence risk in individual patients undergoing surgery with or without radiation therapy for SCCOC. These initial preoperative nomograms are based on a combination of host factors, such as age, sex, and race, and tumor characteristics, such as anatomic subsite affected, clinical estimation of tumor size, and the presence of clinical and/or radiological nodal metastases. These prognostic factors have been previously shown to be predictors of survival and recurrence.[13-18] Our nomograms are able to provide estimates of OS just as the clinical TNM staging system does. In addition, these nomograms are able to provide individualized estimates of risk of cancer recurrence and cancer-specific mortality, outcomes that are not currently reported by the TNM system.

Consider the following 2 patients with cancer of the oral cavity: 1) a 30-year-old otherwise healthy woman with a 2.5-cm tumor in the tongue and a single 1-cm lymph node in the ipsilateral neck; and 2) a 60-year-old male patient who is diabetic, has coronary artery disease, and is an active smoker with a 3-cm hard palate tumor and also a single clinically palpable 2-cm lymph node in the ipsilateral neck. Using the TNM staging system, the oral cavity tumor in both patients would be classified as T2N1M0 (stage III) with an estimated 5-year OS of about 34%. However, this estimation of survival fails to take into account the individual characteristics of the patient or the tumor. Extrapolation of survival estimates for a group of patients with similar tumor characteristics is therefore unreliable when it comes to estimating the same outcome in an individual patient. Nomograms, on the other hand, have the ability to take into account the influence of multiple other factors, and the predicted 5-year OS for these 2 illustrative patients using our nomogram would be 70% and 30%, respectively. Furthermore, when analyzing the distribution of OS probabilities predicted by the nomogram within each TNM stage, a wide heterogeneity is observed (Fig. 7). This observation further reflects the limitations of the TNM system in recognizing differences between individual patients beyond the most basic tumor characteristics.

Figure 7.

Distribution histogram of 5-year predicted overall survival probability for each TNM-AJCC (American Joint Committee on Cancer) stage. These histograms of predicted probabilities illustrate the heterogeneity of the risks within each clinical stage. Without the nomograms, we would consider all patients in each stage as a homogenous group with the same risk; the nomograms help us assign them with different risks.

The nomograms that we have developed have good discrimination ability with good concordance indexes, similar to other widely used nomograms. In our study, we report a lower C-index for recurrence prediction compared to that for OS. The higher number of survival events compared to locoregional recurrence events might explain this phenomenon. As with any predictive model, the individual estimates have a degree of uncertainty, and this uncertainty increases for patients who do not have characteristics similar to those used to generate the model. Validation of our nomograms on an external data set will therefore be important. However, the C-index of our nomograms suggests a sufficient level of accuracy, and is comparable with other published predictive nomograms for sarcoma, breast cancer (NCI Gail Model), and prostate cancer.[9, 10, 19]

Although our nomograms demonstrated good accuracy for prediction of outcomes, there are some limitations to the data and methods that must be considered. For instance, the use of retrospective data to construct the nomograms introduces the risk of treatment selection bias. However, the entire study cohort was treated at a single institution with a uniform treatment policy of primary surgery with or without postoperative radiotherapy. Because the vast majority of patients included in this study were from an era prior to the introduction of postoperative chemoradiation, the influence of this modality is unaccounted for in our analysis. More importantly, the demographic characteristics of the study cohort may be unique to our practice and may not be relevant to risk prediction in other patient populations. External validation of the nomograms in a multi-institutional international data set of patients is therefore an integral part of our plan, and will be the subject of a future communication.

The ability of nomograms to take into account more variables than the conventional TNM staging system also sets up the possibility that a staging system based on nomograms would more accurately stratify patients into groups with similar estimated outcomes than the currently used TNM system.[20] Cancer-related outcomes are influenced by a number of factors, and clinicians commonly consider prognostic factors that are not included in the TNM system for decision-making in actual patient care.[3] The vast majority of these factors have not been incorporated into the staging system because they may not predict outcome “independently” in multivariate models of prognostication. However, it is very clear that outcomes of cancer treatment are influenced by multiple factors. Many of the recognized descriptors of the disease that fail to independently predict prognosis in multivariate statistical analyses may actually work in tandem and have varying degrees of influence on each other, on treatment selection, and on prognosis. Therefore, their exclusion from the staging system and prognostic models could potentially diminish accuracy.

Advances in technology now allow an unlimited number of variables to be combined with the conventional TNM paradigm in order to improve accuracy of the prognostic model.[2, 21] Demographic factors such as age, sex, race, familial history of cancer, socioeconomic status, lifestyles, and habits (tobacco, alcohol, and substance abuse) are only a few of the many factors that impact outcomes of treatment.[13, 14, 22, 23] Clinicians regularly use this information in the decision-making process for treatment of patients with cancer. However, none of the other host characteristics have representation in the current staging system for oral cancer. Medical comorbidity is also a significant determinant of outcome, especially of OS. Patients with significant medical problems may not be capable of receiving treatments equivalent to those of healthier patients. Therefore, comparing outcomes in patients with head and neck cancer without accounting for the influence of medical age, comorbidities or other host factors, introduces significant bias in reporting by considering incorrectly all patients within a particular TNM category or stage as constituting a homogeneous group. Nomograms take into account the heterogeneities of patients with cancer, and therefore calculate survival and risk failure more accurately. Furthermore, nomograms may be used to improve the decision-making of adjuvant treatment and stratification of patients for clinical trials by incorporating factors that are considered in the treatment decision-making process but are not currently included within the current TNM system.

With advances in understanding of cancer biology, we are now faced with a deluge of information on molecular and genetic predictors of prognosis. We now also have an improved understanding of etiopathology of some cancers; human papillomavirus infection in oropharyngeal cancer and Epstein-Barr virus infection in nasopharyngeal cancer are prime examples.[24] Prognostic data of this nature will have to be incorporated into staging systems in order to maintain relevance of any staging system as our understanding of cancer continues to evolve. Nomograms are well suited to adapt to this need, and the nomograms reported in this initial communication are amenable to incorporation in a modular concept (Fig. 8).[3] Such nomograms will enable clinicians to not only predict outcomes for individual patients, but will serve as tools for staging all patients with head and neck cancer.

Figure 8.

A modular prognostic system based on multiple nomograms is shown. From Ref. No. 3 Patel et al., J Surg Oncol. 2008;97:653-657.


We have used a large retrospective database to develop a set of nomograms that can predict 5-year overall survival, recurrence risk, and cancer-specific mortality among individual patients with oral cancer treated primarily with surgery.


No specific funding was disclosed.


The authors made no disclosure.