Introduction
Which of the several models developed to predict outcome following definitive therapy for prostate cancer is the most accurate? An initial analysis of six models showed that the model developed by Thomas Pisansky, MD provided the closest correlation to actual clinical outcome data measuring biochemical freedom from failure (bNED). Following logarithmic transformation analysis, however, all of the models analyzed appeared to be equally predictive of bNED outcome.
The analysis of the models was directed by Benjamin Movsas, MD, Department of Radiation Oncology at Fox Chase Cancer Center. The data were presented at the First Sonoma Conference on Prostate Cancer by Gerald E. Hanks, MD, department chairman.
As we approach the 21st century, clinically useful predictive models are sorely needed to reliably stratify patients for future treatment strategies, Dr. Hanks said. Theyre also very useful in the clinic to help us counsel patients about what to do.
Six Models Were Analyzed
Six models or equations were analyzed in a definitive radiotherapy series of 421 patients with localized prostate cancer. Patients received a median dose of 74 Gy between March 1988 and November 1994. A stepwise Cox proportional hazard multivariate analysis (MVA) was performed using the following covariates: prostate specific antigen (PSA); Gleason score; stage; dose of radiation; 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, defined as two consecutive rises in PSA equaling or exceeding 1.5 ng/mL. The median follow-up was 34 months.
Standard Paradigm Used
Akaikes Information Criteria (AIC) were used to compare the different models. A smaller AIC value corresponds to a statistically more accurate model.
The first model analyzed was the standard paradigm used in the Department of Radiation Oncology at Fox Chase. The three significant factors included in that model are pretreatment PSA, Gleason score, and stage.
The AIC value for the standard model was 817. The AIC values for the other models ranged from 820 for the DAmico and Propert model to 796 for the Pisansky model (Table 1).
Based on the initial analysis, the Pisansky model appeared to be the most predictive due to having the lowest AIC value and the simplicity of the risk estimate, which is the sole predictor of outcome.
It is the only model there that has one degree of freedom. So, from that point of view, its a better model, Dr. Hanks said.
To be useful, Dr. Hanks explained, a predictive model must accurately predict for outcome, in order to provide estimates of risk that are reliable in the patient population that we want to study. And sometimes this is achieved by implementing additional prognostic covariants, or degrees of freedom, to create a better fit.
However, the model that boils down to one risk estimate, or one degree of freedom, provides parsimony that facilitates its application, compared to models that involve multiple degrees of freedom. For example, a model that incorporates three prognostic factors to derive a single risk estimate is advantageous to a model that involves three separate predictive factors.
The differences in accuracy between the models were mostly eliminated by converting them to a logarithmic function, Dr. Hanks said. Following the logarithmic transformation analysis, all of the models appear to be equally predictive of biochemical freedom from failure.
The AIC values differed by only 4 points, with the Pisansky model at 796, the DAmico and Propert models at 800, and all the other models at 799 (Table 1).
Everything in Your Pocket
Summarizing what makes a good model, Dr. Hanks said that it must be clinically useful, needs to be convenient to use in the clinic, to help counsel patients about what their chances truly are, and it needs to be compact. Its best if everything you need to operate the model is in your pocket, when youre in the clinic.