Another substantive advance has been the clarification of prognostic factors for patients with primary American Joint Commission on Cancer stage I cutaneous melanoma ( 1.5 mm) and stage II cutaneous melanoma (>1.51 mm); these factors are predominantly pathologic variables that were described by Clark and colleagues [5], Breslow [6], and other investigators [7]. Survival is routinely predicted almost exclusively on the basis of tumor thickness. Although this is an excellent correlate of prognosis, it is imprecise. Death results from thin melanomas, and survival occurs with thick melanomas. Evaluation of other variables improves the ability to predict survival in patients with melanoma. Variables that can be added include the anatomical site of the primary tumor; sex and age of the patient; mitotic rate of the tumor; presence of ulceration, microscopic satellites, or tumor-infiltrating lymphocytes; tumor regression or angiogenesis; and the Clark level of invasion (maximum penetration of the melanoma lesion into levels of the dermis or subcutaneous tissue) [8, 9].

We previously described a multivariable model that predicted 8-year survival of patients with primary melanomas more accurately than did tumor thickness alone [10]. In that model, six variables were found to independently predict survival. These variables (in order of their relative predictive strength) are mitotic rate, presence of tumor-infiltrating lymphocytes, tumor thickness, site of the primary lesion, sex of the patient, and histologic regression of the tumor. Although this model includes powerful predictors of survival, some of the variables used to generate it are not routinely included in standard pathology reports. Therefore, the model cannot be readily generalized for clinical use.

To document this, we reviewed 100 randomly selected pathology reports of primary melanomas from community-based hospital pathologists, independent pathology laboratories, and specialized dermatopathology laboratories. Tumor thickness was recorded in 76% of cases; Clark level, in 68%; and histologic subtype, in 64%. However, tumor infiltrating lymphocytes and tumor regression were described in fewer than 20% of cases, and the mitotic rate was reported even less frequently (fewer than 5% of cases). Notably, no pathologists specified pure, invasive, radial-growth-phase melanoma, although such lesions almost never metastasize [10]. Given the limited availability of these histologic variables, we sought to develop a prognostic model based on clinical and pathologic data that are routinely available to the clinician. The model is intended to estimate the chance of survival, within 10 years of definitive therapy, in a patient with primary malignant melanoma. The ability to predict outcome more accurately in these patients could identify patients who are at high risk for recurrence; these patients could then be included in trials of adjuvant therapy. Further trials are becoming increasingly important as effective therapies for post-surgical treatment of melanoma are identified and tested [11-13].

Methods

Top Methods Results Discussion Author & Article Info References

Patients

The Pigmented Lesion Group at the University of Pennsylvania prospectively evaluated 624 patients with primary melanoma between 1 September 1972 and 31 December 1979. The pathologic variables were ascertained in an independent study of the primary tumors on two separate occasions by two pathologists who had no knowledge of outcome. Definitive treatment of the primary melanoma consisted of wide re-excision of the primary site to yield negative margins. We excluded 136 cases from analysis. Reasons for exclusion were competing causes of death before 10 years of follow-up with no evidence of melanoma (n = 44), lack of prospective follow-up (n = 22), metastatic disease evident beyond the primary site at presentation (n = 29), inadequate surgical treatment (n = 14), unknown cause of death before 10 years of follow-up (n = 5), occurrence of a high-risk primary tumor after 1979 (n = 2), existence of noncutaneous primary tumors (n = 5), loss to follow-up (n = 10), and unclassified tumor thickness or level (n = 5). Therefore, the final study group consisted of 488 patients who were followed prospectively for at least 10 years. Surviving patients were followed for no more than 20 years. Five patients who died of melanoma after 10 years of follow-up were considered to be 10-year survivors.

Our description of what is generally included in melanoma pathology reports was drawn from the last 100 cases submitted for review at the Hospital of the University of Pennsylvania. Forty-three percent of the reports were from community-based hospitals, 17% were from independent pathology laboratories, and 40% were from specialized dermatopathology practices. We did not intend to evaluate current melanoma pathology reporting in nonacademic settings but rather to evaluate which prognostic factors are frequently reported on pathology reports.

Validation of the model required a test using patient data that had not been used to generate the model. This sample consisted of 142 patients who had primary melanoma and were identified in an identical manner in 1980 and 1981, with blinded histopathologic assessment of melanoma and 10 years of follow-up data.

Clinical and pathologic variables that have been identified as prognostic indicators of survival and that are readily available to the clinician were used to develop the prediction model [7, 9]. Clinical variables included age and sex of the patient and site of the primary lesion. Pathologic variables included histologic subtype, Clark level, and tumor thickness. Tumor thickness was measured from the stratum granulosum epidermidis to the depth of the tumor at its thickest part, according to the method of Breslow [6]. Anatomical site of the primary melanoma was divided into two categories: extremity (upper and lower) and axis. Axial or volar primaries, including melanomas arising on the trunk, head, neck, and palms and soles (volar) and under the nails (subungual), were designated as axis lesions. Lesions on other parts of the body were designated as extremity lesions. Female and male patients were studied. Histologic subtypes of melanoma included superficial spreading melanoma (71%), nodular melanoma (12%), lentigo maligna melanoma (6%), acral-lentiginous melanoma (3%), and other lesions (8%), using standard pathologic criteria. Age at the time of diagnosis was recorded. The Clark level of invasion was determined as described elsewhere [5]. Patient outcome was assigned to two categories: alive at 10 years (with or without evidence of melanoma) or dead from melanoma before 10 years. The 10-year interval was chosen because death from melanoma beyond 10 years is uncommon. We chose to analyze survival as a binary outcome (alive at 10 years compared with dead before 10 years). This method was chosen instead of evaluating the survival time or time to death because our primary objective was to differentiate between patients with high and low probabilities for surviving disease. This differentiation, in turn, can aid patient management and identify candidates for adjuvant therapy trials. Therefore, we were not interested in determining the contribution of prognostic factors that lengthen survival but do not necessarily prevent death.

Statistical Analysis

We used a univariate logistic regression model to test the six clinical and pathologic variables for their association with death. Patient age and tumor thickness were tested as continuous and nominal variables, and the Clark level was tested as a nominal variable. Tumor site was initially evaluated as a variable falling into one of four categories (trunk, head or neck, subungual or volar location, and extremity) but was subsequently reduced to two categories, axis and extremity. Variables that were statistically significantly associated with survival were retained for testing in a multivariable logistic regression model [14]. A manual stepwise procedure was used to determine the best model. The predicted probability that a patient would survive 10 years was generated using the estimated model variables. The logistic equation is presented as a footnote in Table 1.

Figure 1 is a box and whisker plot that shows the distribution of probabilities that were estimated by our model for survival of patients who were alive and those who were dead at 10 years [15]. The boundary lines of the boxes represent the 25th and 75th percentiles of the data; the lines drawn through the interior of the boxes mark the 50th percentiles (the medians). The whiskers are drawn from the edges of each box to the most extreme point a maximum distance of 1.5 times greater and 1.5 times less than the interquartile ranges. Any value more extreme is considered an outlier and is designated by an asterisk. The predictive ability of the four-variable model was compared with that of tumor thickness in two ways. First, receiver-operating characteristic (ROC) curves were calculated [16, 17]. Second, the McNemar test was used to compare the percentages of correct predictions. Goodness of fit was assessed by using the Hosmer-Lemeshow test [14].

View larger version (7K): [in this window] [in a new window]   Figure 1. Comparative boxplots of probabilities that were estimated by the four-variable model for patients who were dead at 10 years and patients who were alive at 10 years. The median value is represented by the line drawn in the interior of the box, and outliers are represented by the asterisks. The proportion of the observed data contained in the box represents the middle 50% of the observation.

Results

Top Methods Results Discussion Author & Article Info References

Overall

Of the 488 prospectively followed patients, 108 died of melanoma before 10 years. The median follow-up was 13.5 years (minimum, 10.0 years; maximum, 20.5 years). However, only the first 10 years of follow-up were used to develop the model. The 10-year survival rate for the entire population was 78%.

Characteristics of Variables That Entered the Model

The median age of patients in the study group was 48 years (range, 4 months to 95 years). In most patients, melanoma was diagnosed when the patients were between 30 and 60 years of age. Four age categories were initially evaluated: younger than 30 years, 30 to 50 years, 51 to 60 years, and older than 60 years. There were slightly more female patients than male patients (54% compared with 46%, respectively); female patients also had a higher survival rate. Our model has two categories for anatomical location: extremity and axis. The former favorably affected survival. Extremity lesions made up 38% of cases, and axial presentations made up the remaining 62%. Tumor thickness ranged from a minimum of 0.07 mm to a maximum of 13.5 mm (mean, 1.63 mm; median, 0.98 mm).

Univariate and Multivariable Analyses

Of the six variables tested using the univariate logistic model, four were independently associated with death (Table 2). In order of their relative predictive strength, these variables were tumor thickness, site of primary tumor (extremity compared with axis), age of the patient ( 60 years compared with >60 years), and sex of the patient. Four thickness categories were tested: less than 0.76 mm, 0.76 mm to 1.69 mm, 1.70 mm to 3.60 mm, and thicker than 3.60 mm; categorization was based on previous studies [10]. Delineation of patient age, tumor site, and tumor thickness into smaller categories did not substantially improve the model; therefore, our categorization of these variables achieved a more parsimonious model. Histologic subtype and the Clark level of invasion were not statistically significant in the multivariable analysis (P > 0.10).

When these four variables were entered simultaneously into the logistic regression model, all remained statistically significant. None of the six two-way interactions was statistically significant (P > 0.2). The adjusted odds ratios and 95% CIs for the four predictor variables are listed in Table 1. The adjusted odds ratio for survival can be interpreted as follows: If a primary melanoma develops on the extremity, the affected patient has 4.4 times the odds of surviving melanoma at 10 years than does a patient with a lesion on the trunk, assuming all other prognostic factors remain constant. Not surprisingly, tumor thickness is the most powerful predictor of death, but the other variables listed also contribute meaningful predictive information. Sex and age of the patient are important prognostic factors even after they are adjusted for thickness and location of the primary lesion. The coefficient and odds ratios for this model may change if other known prognostic factors (such as mitotic rate) are included as predictors.

Generation and Tabular Display of the Model

Through use of the estimated variables from the logistic regression model, point estimates and CIs can be generated for the probability of 10-year survival of patients within the strata of the four predictor variables. Table 3 presents stratification across the four prognostic variables (tumor thickness, site of primary tumor, and patient age and sex) and the corresponding point estimates and CIs for the probability of 10-year survival. For example, the probability of survival of a 45-year-old woman with a 3.2-mm thick melanoma on her right calf can be calculated as follows. Beginning at the far left column, the third thickness category (1.70 mm to 3.60 mm) represents the appropriate tumor thickness; the patient is younger than 60 years of age, has an extremity lesion, and is female. Thus, this patient has a probability of survival of 0.89 (corresponding CI, 0.80 to 0.94) or an 89% chance of surviving melanoma at 10 years.

Comparison with Tumor Thickness as the Sole Predictor of Outcome

The probability for survival determined by using tumor thickness alone is presented in Table 4. First, we compared the predictive ability of the four-variable model with that of tumor thickness alone by comparing the area under the ROC curves. The four-variable model (area, 0.8742 [CI, 0.8432 to 0.9052]) was superior to thickness alone (area, 0.8225 [CI, 0.7845 to 0.8605]) in predicting 10-year survival (P < 0.002).

Second, we evaluated the predictive accuracy of the four-variable model with that of tumor thickness alone by comparing the proportion of correct predictions. Prediction models using logistic regression are often evaluated by establishing a cutoff point; predicted probabilities below the cutoff point are treated as predictors of no event, and predictions at or above the cutoff point are considered to be predictors of the event. A cutoff point of 0.50 is often chosen. However, we felt that a cutoff point of 0.50 was inappropriate for clinical decision making, because values slightly above 0.50 appear too low to predict survival and values slightly below 0.50 appear too high to predict death. Accordingly, we chose 0.70 as the lower limit for predicting survival and 0.30 as the upper limit for predicting death. In other words, a survival probability of 0.70 or greater predicted survival, whereas a survival probability of 0.30 or less predicted death. Values between these limits fall within a gray zone in which the probability was believed to be insufficient to clinically predict the outcome.

Using the 0.30 and 0.70 prediction cutoff points, we calculated the proportion of patients correctly classified by the model. The proportion of correct predictions is the proportion of patients in the total sample whose observed vital status (alive or dead) at 10 years agreed with their predicted vital status. These predictions were compared with those obtained using a logistic regression model with tumor thickness as the only variable. The four-variable model correctly predicted outcome in 74% of cases, and tumor thickness alone was correct in 68% of cases. The model incorrectly predicted outcome in 8% of cases compared with 9% of cases when tumor thickness alone was used. Fewer data are contained in the gray zone when the four-variable model is used than when tumor thickness alone is used (18% compared with 22%; P < 0.04). Overall, a comparison of the estimated probabilities (as paired data) indicates that the four-variable model is more predictive than tumor thickness alone (P < 0.007). The model is particularly useful for predicting death (proportion of dead persons predicted to be dead at 10 years [negative predictive value]). The error rates for the prediction of death were 16% with the four-variable model and 29% with tumor thickness. This represents a reduction of almost 50% in the error rate for the prediction of death.

Figure 1 shows the distribution of the probabilities for patients alive at 10 years and those dead at 10 years. The distribution of estimated probabilities for living patients is generally closer to 1 than is the distribution for dead patients. Some overlap in the distribution is seen in the whiskers and outliers.

Finally, we evaluated the goodness of fit of the models by using the Hosmer-Lemeshow test, which yielded a P value of 0.722. This indicates that the model fits well.

Validation of the Model

Validation of the model required a test using patient data that had not been used to generate the model. Therefore, we used cases of primary melanoma seen by the Pigmented Lesion Group in 1980 and 1981 and ascertained and analyzed these data using the same procedure used with the data from the 1970s. The four clinical and pathologic variables for the validation sample, which comprised 142 patients, are presented in Table 2. The overall survival for this group was 75% (107 of 142 patients). In this cohort, follow-up ranged from 10.1 to 19.7 years (median, 10.8 years).

There was no statistically significant difference in values between the validation sample and the set of patients used to generate the model, except in age (84% of patients were younger than 60 years of age in the validation sample compared with 76% in the set used to generate the model; P = 0.03). For the validation sample, the predictive ability of the four-variable model was again compared with that of thickness alone using the area under the ROC curves. The four-variable model was superior to thickness alone (P < 0.007). We also used the four-variable model to generate predicted survival probabilities for each patient in the validation sample. The proportion of patients correctly classified was 69%, compared with 74% in the original sample used to develop the model.

Discussion

Top Methods Results Discussion Author & Article Info References

During the past several decades, we have learned that melanoma is a heterogeneous disease with wide variation in survival, a finding that reflects several well-defined clinical and pathologic prognostic factors. As originally described by Breslow [6] in 1970, the distance between the top of the epidermis to the base of the melanoma, measured in millimeters by using an ocular micrometer, is a highly accurate, reproducible, and easily obtainable variable that is correlated with the biological risk for metastasis and, therefore, survival. Several reports have confirmed that tumor thickness is an independent variable that can predict outcome in patients with invasive primary melanoma. Over the past two decades, additional prognostic indicators have been identified, including patient sex and age, site of lesion, presence of ulceration, presence of infiltrating lymphocytes, and number of mitoses. The difference in survival between men and women has not been explained. Men are more likely to have axis lesions than extremity lesions in comparison with women, but this only partly explains why the male sex carries an inherent disadvantage in terms of survival from melanoma.

We previously developed a multivariable prognostic model that predicts survival of patients with primary melanoma more accurately than does tumor thickness alone [10]. That model first required distinguishing nontumorigenic, invasive, radial growth-phase lesions (those incapable of metastasis) from lesions with a tumorigenic (vertical) growth phase. For lesions with a vertical growth phase, six variables were found to be independent predictors of survival. However, three of these variables—mitotic rate, presence of tumor-infiltrating lymphocytes, and evidence of histologic regression—are rarely reported on standard pathology reports. Furthermore, pure radial growth-phase lesions are not routinely separated from those in the vertical growth phase. For these reasons, the multivariable prognostic model is not applicable to many patients with melanoma. Therefore, we developed an alternative, four-variable prognostic model that uses data readily available to the clinician.

For certain patients with melanoma, determining prognosis is not difficult. Patients with very thin melanomas (<0.76 mm) have an overall chance of survival that approaches 96%. In addition, patients with very thick primary lesions, clinical involvement of multiple regional lymph nodes, or distant metastatic disease have a greater than 90% chance of dying of melanoma by 10 years. Our four-variable model cannot improve the accuracy of predicting survival or death in these patients, but for the remaining patients (particularly those with primary lesions of intermediate thickness) the outcome is less certain. Although this analysis confirms the importance of tumor thickness in predicting 10-year survival in patients with invasive primary cutaneous melanoma, survival can be predicted more accurately by using additional prognostic information that is readily available. Several clinical situations may make particular use of the four-variable model. For melanoma lesions of intermediate thickness (1.70 mm to 3.60 mm), the thickness model predicts a probability of survival of 0.59; probabilities of survival with lesions of this size that are derived from the four-variable model range from 0.89 to 0.24, depending on the sex and age of the patient and the location of the primary lesion. A woman with a lesion on an extremity has a much better probability of survival when the four-variable model is used (0.89 to 0.73, depending on the age of the patient) than when tumor thickness alone is used (0.59) (Table 3). Similarly, men with relatively thin melanomas (0.77 mm to 1.69 mm) located on the trunk have a less-favorable outcome using the four-variable model (0.75 to 0.50) than when using tumor thickness as the sole prognostic factor (0.83).

Clearly, our model has some limitations. Additional prognostic factors, some known and some unknown, play an important role in determining survival from melanoma. In addition, we validated this model using patients from the same institution in which the model was developed. Further corroboration of this model using a different patient population from a different setting is an important step in evaluating the model's validity and usefulness.

Physicians caring for patients with melanoma are frequently asked questions about prognosis. Answers to these questions may affect patients' quality of life and personal plans. Therefore, the physician must carefully distinguish prediction of the outcome of a population of patients from that of an individual patient. Prognostic models greatly affect oncologic decisions about the management of patients, particularly in the adjuvant setting. This four-variable model provides clinical investigators with a practical tool that more accurately classifies risk in various subsets of patients with melanoma than do previous models; it may be used to define high-risk patients for adjuvant clinical trials. The model will also be useful for stratifying patients in clinical trials, thereby ensuring that important prognostic factors will be well balanced between treatment groups. As with other malignant conditions, optimal management and follow-up of patients with primary cutaneous melanoma require an understanding of the factors associated with prognosis.

Dr. Schultz and Ms. Synnestvedt: University of Pennsylvania, 528 Blockley Hall, Biostatistics Unit, 428 Guardian Drive, Philadelphia, PA 19104.

Dr. Trock: Lombardi Cancer Center, 2233 Wisconsin Avenue NW, #535, Washington, DC 20007.

Dr. Elenitsas and Halpern: Hospital of the University of Pennsylvania, Department of Dermatology, 3400 Spruce Street, Philadelphia, PA 19104.

Dr. Elder: Hospital of the University of Pennsylvania, Department of Pathology, 3400 Spruce Street, Philadelphia, PA 19104.

Dr. Clark: Beth Israel Hospital, Department of Pathology, Boston, MA 02215.