A Simple and Accurate Model for Prediction of Cancer-Specific Mortality in Patients Treated with Surgery for Primary Penile Squamous Cell Carcinoma

Carcinoma of the penis is a rare solid tumor with an estimated incidence of 7 to 9 per million in the United States of America (1–3). Its incidence is similar in Europe (1 to 9 per million; ref. 4). Conversely, the incidence is exponentially higher in developing nations (190 per million; ref. 5). In North America, 95% of all penile cancer are squamous cell carcinoma (1). Low incidence of primary penile squamous cell carcinoma (PPSCC) in North America makes it difficult to provide accurate and valid estimates of prognosis after treatment of the primary penile tumor (6–10). To address the void in the ability to predict the prognostic of patients with PPSCC, investigators from Italy developed a model for prediction of cancer-specific mortality (CSM) after treatment of PPSCC (11). The Italian model relies on 8 predictor variables. These include histologic grade, in addition to 7 other variables (tumor thickness, stage, growth pattern, presence of tumor vessel embolization, lymph node status, infiltration of the corpora cavernosa, infiltration of the corpus spongiosum, and urethral infiltration). Unfortunately, several of the predictor variables required by the nomogram such as tumor thickness, growth pattern, and embolization group are not available in routine pathologic reports. This limitation undermines the applicability of this nomogram. Moreover, the Italian nomogram cannot be applied to patients treated with interstitial brachytherapy, where the prediction variables required in the Italian nomogram are not available (12). Finally, the Italian nomogram may not be applicable to patients with PPSCC treated in the United States due to potential differences that may exist in the diagnosis, stage, and treatment of patients with PPSCC between Italy and the United States. To circumvent these problems, we attempted to develop a simpler and at least equally accurate predictive rule using a large population-based cohort from the United States, to assist clinicians with the prediction of prognosis in patients with PPSCC.

Patients diagnosed with PPSCC between 1988 and 2004 were identified within 9 Surveillance, Epidemiology and End Results (SEER) registries (13). These included the Atlanta, Detroit, San Francisco-Oakland, Seattle-Puget Sound metropolitan areas, as well as the states of Connecticut, Hawaii, Iowa, New Mexico, and Utah. The characteristics of the SEER population are comparable with the general population of the United States and represent a 10% sample (13). The penile carcinoma diagnostic code [International classification of Disease for Oncology, Second edition for penile lesions (C60.X)] was used as the main inclusion criterion. Among those, only patients with squamous cell histology were considered. Exclusions consisted of patients who underwent surgeries other than excisional biopsy, partial or total penectomy, and of unknown disease stage or unknown histologic grade. The tumor stage was defined according to the SEER stage classification (localized versus regional versus metastatic). Due to low number of PPSCC grade IV patients, pathologic grade was divided into three categories (I versus II versus III-IV). The cause of death (cancer specific versus noncancer related) was defined according to the SEER mortality cause assignment. Patients who did not die of PPSCC were considered as having succumbed to other causes.

Statistical analyses. The cohort was used to fit univariable and multivariable competing risks regression models according to the criteria of Fine and Gray (14). Competing risks regression models account for the effect of noncancer related mortality. Strong effect of competing mortality may result in extensive censoring due to cancer-unrelated deaths. Censoring due to noncancer-related mortality may artificially reduce the pool of individuals at risk of PPSCC-specific events. This may in turn overestimate the effect of PPSCC-specific mortality. The predictors consisted of patient age, race, SEER stage (localized versus regional versus metastatic), histologic grade (I to IV), pathologic lymph node stage (pN₀ versus pN_1-3 versus pN_x), and of the year of diagnosis. Stepwise variable removal was then applied to the full multivariable model, according to the Akaike's information criterion, with the intent of developing the most accurate and parsimonious model (15, 16).

Actuarial survival probabilities were estimated using the Kaplan-Meier method. Subsequently, the multivariable Cox regression coefficients of the most accurate and parsimonious model were used to generate a look-up table predicting individualized CSM probabilities at 5 y after surgical removal of PPSCC.

To maximize the efficiency of the data, the accuracy of the predicted 5-y CSM rates was assessed internally, using the Harrell's concordance index (15, 16). The leave-one-out internal validation was then done to quantify the most unbiased accuracy estimate of the model, where accuracy was defined as the ability of the model to discriminate between two patients with different survival lengths. Finally, the relationship between predicted and observed CSM rates was explored graphically using the val.surv function from the R statistical package. The model-predicted probabilities of 5-y CSM were plotted against the observed 5-y CSM rates (15, 16). The resulting plot was used to assess model calibration. Perfect calibration consists of a 1:1 relationship between predicted and observed values. Subsequently, we repeated all analyses to specifically quantify the added value of pathologic nodal stage in the entire cohort, as well as in patients who underwent a formal lymph node dissection. All statistical tests were done with S-Plus Professional and R statistical package. Statistical significance was set at 0.05.

Overall, 1,151 patients were treated for PPSCC within the 9 SEER registries during the study period. Inclusion and exclusion criteria are shown in Fig. 1. Of the 856 assessable patients with PPSCC, 178 (20.8%), 646 (75.5%), and 32 (3.7%). patients were respectively treated with an excisional biopsy, a partial, or a radical penectomy (Table 1). According to SEER stage, localized PPSCC was recorded in 551 patients (64.4%) versus regional in 277 (32.3%) and metastatic in 28 patients (3.3%). Grades were divided between 276 grade I (32.4%), 387 Grade II (45.2%), and 193 Grade III to IV (22.5%). Lymph node dissection was done in 139 patients (16.2%). Pathologic N stages were as follows: pN₀ in 57 patients (6.6%) versus pN_1-3 in 82 patients (9.6%) versus pN_x in 717 patients (83.8%).

Figure 2A shows the cumulative incidence of cancer-specific and other-cause mortality after surgical removal of PPSCC. At 5 years, 16.8% died of PPSCC versus 28.7% died from other causes. Figure 2B shows the CSM rate according to the Kaplan-Meier method. At 5 years, 17.7% died of SCC, after censoring for other cause of mortality and for loss to follow-up. Stratified Kaplan-Meier plots illustrate the effect of stage and grade on the actuarial CSM rates (Fig. 2C-D).

Table 2A lists the univariable and multivariate competing-risks regression models fitted in the entire cohort. In univariable competing-risks regression analyses, SEER stage and histologic grade represented statistically significant predictors of CSM (all P < 0.001). In the full multivariable model, only stage (P < 0.001) and grade (P < 0.001) achieved independent predictor status.

Table 2B lists the univariable and multivariate Cox survival models fitted in the entire cohort. In univariable analyses, SEER stage and histologic grade represented statistically significant predictors of CSM (all P < 0.001). In the full multivariable model, only stage (P < 0.001) and grade (P < 0.001) achieved independent predictor status.

Table 3 shows a look-up table defined using the regression coefficients of the variables included in the final model (stage and grade). This look-up table predicts the individual probability of 5-year CSM after surgical removal of PPSCC. The accuracy of the look-up table was 73.8% in the internal validation cohort of 856 patients.

Figure 3 shows the graphical comparison between the look-up table predicted probabilities of 5-year CSM and the actual fraction surviving. The curve virtually follows the 45-degree slope, which indicates virtually ideal performance, for predicted CSM risk value of 45% or less. The paucity of data points above 45% predicted risk resulted in graphically less well-balanced ratio between predicted and observed rates.

Analyses addressing the effect of nodal stage (N₀ versus N_1-3 versus N_X) showed that these 3 N stage categories failed to show independent predictor status (P = 0.2) and did not significantly improve the model accuracy. Actually, accuracy decreased from 73.8% to 73.7%. Finally, analyses done in the subset of patients with pathologically staged lymph nodes revealed independent predictor status for lymph node stage (P = 0.045).

CSM after surgical removal of PPSCC may be highly variable (2, 17–21). Stage and grade represent established predictors of CSM (11, 20, 22, 23). However, until recently, no prognostic model allowed to rely on the combined effects of these two predictors to provide individual estimates of CSM after removal of PPSCC. To address this void, investigators from Italy developed a nomogram predicting PPSCC-specific mortality (11). This model represents the first standardized and individualized prognostic tool for patients treated with partial or total penectomy. Unfortunately, its use is rarely possible because besides stage and grade, the Italian nomogram requires six other variables that are not routinely reported after penectomy. For example, tumor thickness, growth pattern, and vessel embolization are all required for prediction of PPSCC-specific mortality but are mostly unavailable in standard pathology reports. In consequence, the Italian nomogram is not applicable to a large number of patients treated for PPSCC. Moreover, the nomogram is not applicable to an equally large number of patients treated with organ preservation in the form of interstitial brachytherapy, in whom pathologic detail required by the Italian nomogram may not be available (12). Finally, the Italian nomogram may not be applicable to patients from outside of Italy, for example from the United States, because the natural history, the diagnosis, staging, and treatments may differ between different countries. To circumvent these limitations, we attempted to develop a simpler prognostic rule that could be used in men from the United States. Our goal was to rely on routinely available pathologic variables and to achieve an equal accuracy, relative to the previously reported Italian nomogram (11).

Our analyses relied on the SEER database that represents 10% of the United States population and is considered as the largest cancer registry (13). A total of 856 PPSCC patients were identified within the SEER database, which exceeds the sample size of all previous analyses, where populations ranged from 32 to 700 (3, 11, 24, 25). Our prognostic rule exclusively relied on stage and grade and achieved 73.8% accuracy (defined as discrimination between two individuals with different survival times) in predicting 5-year CSM. This accuracy figure is at least equal to the Italian nomogram (72.8% accuracy), which relied on tumor thickness, growth pattern, presence of tumor vessel embolization, in addition to stage and pathologic grade (11). In general, the predictions of our prognostic rule very closely approximated the observed rate of CSM. This characteristic was used to assess the calibration of the model. The current prognostic rule may therefore more accurately risk stratify patients with PPSCC than the use of stage or grade separately. Such risk stratification may prove useful in the context of adjuvant therapy considerations, as well as in the context of future clinical trials (5, 12, 26, 27). Moreover, the predictions may assist clinicians with the determination of the frequency and the extent of follow-up.

Although, the T, N, and M stage assignment was done using a mix of clinical and pathologic diagnoses, our model achieved higher accuracy than the previously reported Italian model that relied on detailed pathologic information, and this applied to all stage categories, including patients who did not undergo a formal lymph node dissection. This observation emphasizes the importance of clinical staging and indicates that clinical staging in combination with the grade of the primary tumor may provide the clinician with an excellent ability to predict the probability of CSM. It is conceivable that the inclusion of detailed pathologic information could further enhance the accuracy of the current model.

Our prognostic rule, such as the Italian nomogram, is based on Cox regression modeling. We have shown that a substantial proportion of patients with PPSCC may die of noncancer related causes (Fig. 2A). In such setting, Cox regression models may overestimate the true rate of CSM and competing-risk regression methods, such as developed by Fine and Gray (14), may provide more unbiased CSM estimates. To address this potential limitation of Cox regression modeling, we compared the CSM rates according to both methods: Cox and competing risks regression. The Kaplan-Meier based predictions of CSM that we used in Cox regression models were virtually the same as cumulative incidence estimates that are used in competing-risks regression model (Fig. 2A-B). For example at 5 years, the CSM rates were 17.7% using the Kaplan-Meier method versus 16.8% using the competing-risks method. Moreover, we confirmed the independent predictor status of the two key predictors (stage and grade) in Cox regression, as well as in competing-risks regression models (Table 2A and B). Based on the striking similarity of the results obtained with the two different modeling techniques, we relied on Cox regression modeling for the development of our prediction rule. Our decision was based on exclusive use of Kaplan-Meier and/or Cox regression techniques in previous analyses focusing on PPSCC. The use of the same methodology in the current analysis allows valid direct comparisons.

The simplicity of our prognostic rule represents a strength. Unlike models aimed at identification of risk factors, where an extensive number of candidates and potential confounders are examined, predictive and prognostic models should focus on accuracy and parsimony. In clinical practice, extensively complex models, where a large number of predictors are used may not be implemented or may be abandoned in favor of simpler models. Time considerations and potential unavailability of prediction variables represent important reasons for lack of clinical implementation of complex tools. Our prognostic rule is devoid of these limitations as it rests on two key predictors that are invariably available in the charts of patients with SCC of the penis.

The weakness of our prognostic rule consists of lack of central pathology. However, the population-based nature of the cohort of our prognostic rule compensates for this limitation. Moreover, central pathology is unavailable in routine clinical practice and the previous nomogram also did not benefit of central pathology. Besides stage and grade, other variables may predict CSM. Smoking status represents a well-recognized predictor of CSM (23, 28–32). Unfortunately this information could not be obtained from the SEER database. Other similar variable include neonatal circumcision status, HIV, and human papillomavirus infection status (23, 28–32).

Despite its limitations, our prognostic rule represents the simplest, yet the most accurate method for prediction of CSM in patients with PPSCC diagnosis and treatment in the United States.

No potential conflicts of interest were disclosed.