We commend the efforts of Dr. D’Amico in developing and presenting his stratification schema of prognostic variables for patients with localized prostate cancer. Clearly, the benefit of such a stratification scheme is to better identify those who may be eligible for either more or less aggressive therapy, as predicted by this study. His stratification schemes have been demonstrated to be effective.
Nevertheless, we have concerns about the scheme that Dr. D’Amico has devised and applied. The three groups were constructed logically, using reasonable cutoff points for stage, grade, and prostate-specific antigen (PSA) level. However, it is not clear whether they represent the optimal grouping for prognostic discrimination. Shipley et al developed an empiric risk-stratification scheme by analyzing pretreatment variables and outcome results in a large multi-institutional cohort of patients treated with three-dimensional (3D) conformal radiotherapy (CRT). It is not clear that Dr. D’Amico’s scheme, although easier to remember, is as accurate as Dr. Shipley’s.
Another approach to risk estimation is the use of prognostic nomograms that compute continuous probabilities of an outcome. Rather than placing the patient into one of a few risk groups, prognostic nomograms estimate probabilities, usually of freedom from progression, although other end points, such as survival or disease-specific survival, can be predicted as well.
Such nomograms have been developed for prostate cancer treated with surgery, 3D CRT, and brachytherapy (Figures 1, 2, and 3).[2-4] Their advantage over risk-stratification prediction systems is likely associated with the use of continuous-risk scales for relevant technique parameters, rather than condensing sections of the risk spectra into heterogeneous risk groups. Due to improved accuracy, these nomograms may benefit patient counseling, and more accurate prediction should translate into better treatment decisions.
We compared our radiotherapy nomogram to several published risk stratification schemes and found that the nomogram provided better discrimination when applied to a validation dataset of 932 patients. To illustrate the use of this nomogram and compare its prediction with risk stratification, consider a patient with clinical stage T2c disease, a pretreatment PSA level of 6 ng/mL, a biopsy Gleason score of 9, and a planned radiation dose of 66.6 Gy without neoadjuvant hormonal therapy. Using a common risk-stratification scheme (without the percentage of core biopsy information), this patient would be classified into the most favorable group, which has an associated freedom-from-recurrence prediction of 81% at 5 years.
According to the external radiotherapy nomogram, the same patient’s predicted probability of 5-year freedom from recurrence is 24%. Using the preoperative surgery nomogram, the predicted probability of 5-year freedom from recurrence is 68%, and with the brachytherapy nomogram, it is 62%. Thus, risk stratification may influence this patient to choose radiation therapy over surgery or brachytherapy, although with the nomogram approach, this particular patient would seem to have a better chance of achieving freedom from progression with surgery or brachytherapy rather than low-dose radiation.
Risk Stratification vs Nomogram Predictions
The risk-stratification approach potentially produces heterogeneous strata with respect to nomogram predictions and, assuming the nomogram is more accurate, could be misleading when deciding on a treatment strategy. The latter possibility is illustrated by the number of patients with predictions of low freedom from recurrence by the nomogram despite "favorable" risk stratification, and the number of patients with predictions of high freedom from recurrence by the nomogram despite "unfavorable" risk stratification (see Figure 4). Thus, risk stratification may not always offer the patient the best estimate of treatment efficacy and may, in fact, lead to a poor treatment choice.
In the development of the three nomograms, the pretreatment PSA value, Gleason grade, and clinical stage have been identified as important risk variables in patients with prostate cancer. It appears that the percentage of positive core biopsies provides additional prognostic information, beyond what is conveyed by the three standard predictor variables as combined by Dr. D’Amico. But if his stratification scheme is not optimal, the additional benefit of biopsy information remains uncertain. One would want to demonstrate such benefit as an addition to an optimal risk assessment, and we believe nomograms offer a more accurate initial approach. In fact, we recently assessed the prognostic value of adding biopsy results to nomogram predictions of progression after radical prostatectomy and showed that adding the percentage of positive cores provides limited benefit.
Quantitative analysis of biopsy results generates considerable data in addition to the percentage of positive cores. Other readily available systematic biopsy features also appear to yield valuable information. We recently used a database of 1,047 patients treated with permanent brachytherapy to examine 26 pretreatment factors that included the following:
Clinical stage, biopsy Gleason sum, and pretreatment PSA level
Mean percentage of cancer in an involved core
Maximum percentage of cancer
Mean primary Gleason grade
Mean secondary Gleason grade
Maximum Gleason grade (primary or secondary)
Percentage of cancer in the apex, midregion, and base
Percentage of positive cores
Maximum primary Gleason grade in the apex, midregion, and base
Maximum secondary Gleason grade in the apex, midregion, and base
Maximum percentage of cancer in the apex, midregion, and base
Maximum Gleason grade in the apex, midregion, and base
Maximum primary Gleason grade
Maximum secondary Gleason grade
Four modeling strategies were compared. As a base or first model, we considered the pretreatment PSA level, clinical stage, and biopsy Gleason sum as predictors. For the second model, we added the percentage of positive cores. The third modeling strategy was to use stepwise variable selection of only those factors (from the pool of 26 variables) that were statistically significant. The fourth strategy was to apply principal components analysis, which has theoretical advantages over the other strategies. Principal components analysis creates component scores that account for maximum variance in the predictors.
An analysis of the four models showed substantial prediction differences. Ranking each model based on its ability to predict freedom from recurrence, we determined that the base modelmodel 1, using pretreatment PSA level, clinical stage, and Gleason sumhad the lowest predictive value. The next best model added percentage of positive cores. The variable reduction approach via principal components analysis produced the most accurately predictive model, which could easily be implemented with standard database software or a handheld device. Also, principal components analysis is advantageous over stepwise variable selection because less information content is lost, and the resulting predictor variables do not have inflated effects.
This novel approach of variable reduction modeling can be incorporated into the nomogram modeling techniques described. In addition, this method may provide a useful extension to the important models developed by Dr. D’Amico.