A prognostic model for predicting 10-year survival in patients with primary melanoma

Authors: Schuchter L, Schultz DJ, Synnestvedt M, Trock BJ, Guerry D, et al.
Source: Annals of Internal Medicine. 125:369-375. September 1, 1996.
Institutions: University of Pennsylvania; Lombardi Cancer Center, Washington DC; Beth Israel Hospital, Boston.
Financial support: None indicated.

This article describes the development of a predictive model to help estimate disease prognosis, using a number of clinical variables. An accompanying "Perspective" piece gives a good overview of how such models can be used in clinical practice.



Survival with primary cutaneous melanoma is routinely predicted based on tumor thickness, but this method is imperfect. The authors previously described a six-variable model that was a better predictor of survival but which used a number of pathologic variables frequently not available to the clinician. In this paper, they describe the development of a model that incorporates data routinely available, thus with wider clinical applicability.


The outcome of interest was 10-year survival after surgical treatment of primary melanoma. Patients were classified as either alive at 10 years (with or without melanoma) or dead before 10 years from melanoma (patients dying before 10 years follow-up from non-melanoma causes were excluded).


The 10-year survival rate (excluding deaths from non-melanoma cause) was 78%.

Perspective -- Predicting clinical states in individual patients

Leonard E. Braitman, PhD, and Frank Davidoff, MD

This article reviews the basic concepts behind probability modelling and its application to clinical medicine. Before such a model can be applied to a given patient, two basic issues must be adressed. First, how good is the model in general? Second, how applicable is the model to the specific patient.

The first issue involves the quality of the model. How well does the model "fit" the patients that were used to develop it? How does it compare to other models? Was it tested on patients other than those used to develop it?

The second issue requires knowledge of the clinical situation, of the specific patient. Is the patient sufficiently similar to those for whom the model was developed? Does the model provide the type of information we are looking for?

In the Perspective piece, the authors deal with these issues in detail, adding much insight to the process. I would strongly recommend reading this piece.


Although the model presented here appears to be more precise than the tumor-thickness-alone model, the advantage is surprisingly small. This may be a case where statistically significant does not mean clinically very significant.

The authors state that their model adds the most information for patients for whom the thickness-alone model predicts an intermediate probability of survival, and where their model yields either a higher or lower estimate (thickness alone model yields a survival probability of 0.59; their model yields probabilities ranging from 0.24 to 0.89). They do not indicate how accurate their model is for these intermediate cases, however. Without knowing this, we cannot judge whether or not the four-variable model is really more precise in this subgroup of patients.

Prognostic models such as the one described here can help clinicians and patients make more informed choices about approaches to therapy. In addition, such models can assist in assessing the results of newer therapies, by comparing predicted survival to achieved survival. In randomized trials, they can help insure that groups are prognostically "comparable". With the increasing emphasis on hard data in health care, prognostic models will be used to generate numbers for various purposes (risk-adjustment in capitated systems, for example). When applying probabilistic models to individual patients, the various considerations detailed in the "Perspective" article should be kept in mind.

A note on receiver operating curves

Most diagnostic tests involve criteria that can be varied. As these criteria are varied, there is usually a tradeoff between sensitivity and specificity. An example will make this clearer. If fasting blood sugar is used as a test for diabetes and if the cutoff for abnormal is set at 200 mg/dl, the test will be highly specific (very few false positives), but sensitivity will be very poor (many diabetics will be missed, there will be a lot of false negatives). As the cutoff point is gradually lowered, the specificity will decrease while the sensitivity increases. At a cutoff of 100, the test will be extremely sensitive (picking up nearly all diabetics) but specificity will be terrible. For each possible cutoff point, we can determine the test's specificity and sensitivity. We can then plot sensitivity versus specificity (actually, what is usually plotted is sensitivity vs. 1-specificity). This sensitivity vs. specificity curve is a receiver operating curve. The less of a tradeoff there is between sensitivity and specificity, the better the test is. This is tantamount to saying: the closer the area under the ROC is to one, the better the test is.

The figure below represents two typical ROC curves. Curve B represents less of a tradeoff between the sensitivity and the specificity than curve A, has an area under the curve closer to one, and is the "better" test.


Reader comments

Submit a comment about this article


Full text of this article and of the Perspective piece from the ACP website.

References related to this article from the NLM's PubMed database.

Return to Journal Club home page

Site Meter