Background: As we approach the 21st century, clinically useful predictive models for prostate carcinoma are urgently needed to stratify patients reliably for future treatment strategies. Recently, many investigators have developed models that employ prostate specific antigen (PSA)-based constructs or groupings in an attempt to predict outcome accurately following definitive radiotherapy. This investigation was conducted to determine which of these models provides the closest "fit" to independent clinical outcome data measuring biochemical freedom from failure (bNED control), thereby warranting further exploration.
Methods: Six models were analyzed in a definitive radiotherapy series of 421 patients with localized prostate carcinoma treated with a median dose of 74 Gray (Gy) between March 1988 and November 1994. A stepwise Cox proportional hazards multivariate analysis (MVA) was performed to predict for bNED control using the following covariates: PSA, Gleason's score, stage, dose, PSA density, and perineural invasion. Subsequent MVAs were performed for each model incorporating the new construct or prognostic groupings. The adequacy of the models was confirmed using plots of score residuals against time to bNED failure and comparisons were made used Akaike's Information Criteria (AIC) in which a smaller value corresponds to a statistically improved model based on explained variation and the number of predictors. Because PSA was distributed in a log-normal fashion in the current study population, the model-building process was duplicated using a logarithmic transformation analysis. Biochemical failure was defined as 2 consecutive elevations in the PSA > or = 1.5 ng/mL. The median follow-up time was 34 months (range, 2-87 months).
Results: Initially, the model developed by Pisansky et al. appeared the most predictive due to the parsimony in their risk estimate, which is the sole predictor of outcome, as well as its associated lowest AIC value. However, after the logarithmic transformation analysis, all the models appeared to be equally predictive of bNED outcome.
Conclusions: A plethora of accurate models for predicting outcome following definitive radiotherapy for prostate carcinoma recently have been engineered, all of which are essentially equally predictive in this data base (via a logarithmic conversion process). This analysis should be corroborated in other large radiotherapy series.