Decision curve analysis to compare 3 versions of Partin Tables to predict final pathologic stage

Abstract

Objective

To perform a decision curve analysis (DCA) to compare the Partin Tables 1997, 2001, and 2007 for their clinical applicability.

Material and methods

Clinical and pathologic data of 687 consecutive patients treated with open radical prostatectomy for clinically localized prostate cancer between 2003 and 2008 at a single institution were used. DCA quantified the net benefit relating to specific threshold probabilities of extraprostatic extension (EPE), seminal vesicle involvement (SVI), and lymph node involvement (LNI).

Results

Overall, EPE, SVI, and LNI were recorded in 17.8, 6.0, and 1.2%, respectively. For EPE predictions, the DCA favored the 2007 version vs. 1997 for SVI vs. none of the versions for LNI.

Conclusions

DCA indicate that for very low prevalence conditions such as LNI (1.2%), decision models are not useful. For low prevalence rates such as SVI, the use of different versions of the Partin Tables does not translate into meaningful net gains differences. Finally, for intermediate prevalence conditions such as EPE (18%), despite apparent performance differences, the net benefit differences were also marginal. In consequence, the current analysis could not confirm an important benefit from the use of the Partin Tables and it could not identify a clearly better version of any of the 3 available iterations.

Introduction

Prostate cancer is one of the most common male malignancies in the Western world. Unfortunately, treatment decisions are hampered by lack of standardization. Heterogeneity of clinical tumor characteristics represents one of the complexities that add to difficulties related to clinical decision-making [1]. To facilitate clinical decision-making, Partin et al. pioneered standardized predictions of pathologic stage prior to therapy [2]. In series, their efforts resulted in 3 generations of the Partin Tables [3], [4], [5]. Those valuable tools were validated in North America and Europe [6], [7]. However, recently certain limitations were recorded in European patient cohorts. For example, in men from Germany, France, and Italy, the 2007 version of the Partin Tables showed suboptimal performance characteristics [8], [9]. This observation prompted us to examine the 1997 and 2001 versions of the Partin Tables with the intent of comparing them with the 2007 version. Commonly, prediction models are evaluated by comparing the predictions of the model with the actual patient outcome. Ideally, these models should not only be subject to internal validation, but also be tested by an external data set. The results of these validations are usually expressed as area under the receiver operating characteristic curve (AUC). The AUC quantifies the model's ability to discriminate between patients with and without a specific feature. As such, the AUC is commonly used to compare prediction tools [10]. However, the AUC centers solely on the accuracy of a prediction tool; it does not address the clinical value of a model [11]. Decision curve analysis (DCA) represents a new analytical technique, which incorporates clinical consequences of a decision [11]. Hereby, DCA estimates the net benefit of a statistical tool for a given clinical threshold probability. Subsequently, it allows the determination whether a model is clinically useful or not. This methodology relies on a single metric that is applicable to all predictive models.

To the best of our knowledge, the 3 versions of the Partin Tables were never subjected to DCA comparison within the same patient population. We relied on DCA to quantify the potential gains related to the use of one version of the Partin Tables relative to another.