|Home | About | Journals | Submit | Contact Us | Français|
Prostate cancer is a very complex disease, and the decision-making process requires the clinician to balance clinical benefits, life expectancy, comorbidities, and potential treatment related side effects. Accurate prediction of clinical outcomes may help in the difficult process of making decisions related to prostate cancer. In this review, we discuss attributes of predictive tools and systematically review those available for prostate cancer. Types of tools include probability formulas, look-up and propensity scoring tables, risk-class stratification prediction tools, classification and regression tree analysis, nomograms, and artificial neural networks. Criteria to evaluate tools include discrimination, calibration, generalizability, level of complexity, decision analysis, and ability to account for competing risks and conditional probabilities. We describe the available predictive tools and their features, focusing on nomograms. While some tools are well-calibrated, few have been externally validated or directly compared to other tools. In addition, the clinical consequences of applying predictive tools need thorough assessment. Nevertheless, predictive tools can facilitate medical decision-making by showing patients tailored predictions of their outcomes with various alternatives. Additionally, accurate tools may improve clinical trial design.
In the Unites States, prostate cancer (PCa) is the most common solid malignancy and the second leading cause of cancer death in men . PCa is a very complex disease, and the decision-making process at the basis of the management of radical treatments is challenging because it requires a fine balance among expected clinical benefit, life expectancy, comorbidities, and potential treatment-related adverse events. The prediction of clinical outcomes is therefore critical. Accurate estimates of stage and of the likelihood of cancer diagnosis, clinical significance, treatment success, complications, and long-term morbidity are important for patient counseling and informed decision-making. By understanding the most probable endpoint of a patient’s clinical course, physicians may modify treatment and post-treatment strategies in order balance benefits and adverse events of treatment. Prediction also allows patients to choose responsibly among the different treatment strategies proposed by the clinicians. Properly informing the patient of these likelihoods could improve his satisfaction after treatment. Lack of patient involvement in treatment decision has been identified as a major risk factor for regret of treatment choice [2, 3]. Therefore, accurate estimates of risk help physicians make specific management recommendations.
Traditionally, physician judgment has formed the basis for risk estimation, patient counseling, and decision making. However, clinicians’ estimates are often biased due to both subjective and objective confounders [4–7]. To obviate this problem and to obtain more accurate predictions, researchers have developed predictive tools that are based on statistical techniques . Recently, predictive tools have been introduced to predict the outcome of interest for the individual patient. Predictive tools have been shown to perform better than clinical judgment when predicting probabilities of outcome [7, 9, 10]. That said, physician input is obviously essential in medical decision-making, both for the measurement of variables that are used in the prediction process and for the interpretation and application of prediction tool-derived outcome predictions in clinical practice.
There are a number of types of prediction tools such as Kattan-type nomograms [11, 12], risk groupings [13, 14–16], artificial neural networks (ANNs) , probability tables [18, 19], classification and regression tree (CART) analyses [20, 21], probability formulas, look-up and propensity scoring tables, and risk-class stratification prediction tools. In this review, we describe criteria for evaluation of predictive tools and present an inventory of available predictive tools available for PCa, which were identified in a search of MEDLINE literature from January 1966 to April 2009. For each of the tools, we describe the patient population, the outcomes predicted, and individual tool characteristics such as predictor variables, discrimination estimates, and whether the tool has been internally and/or externally validated.
Prediction tools can be compared based on several criteria: discrimination, calibration (correlation between predicted and observed risk throughout the entire range of predictions), generalizability, level of complexity, adjustment for the effect of competing risks, use of conditional probabilities, and application of decision-analysis. The most important of these considerations are discrimination and calibration.
A good predictive tool is able to discriminate between patients with or without the outcome of interest. Discrimination is quantified using the area under the curve (AUC) for binary outcomes (e.g., presence or absence of cancer on biopsy), the c-index for censored data (e.g. recurrence after radical prostatectomy), and the Brier score . For both AUC and c-index, 0.5 represents no discriminating ability (coin flip), whereas a value of 1.0 represents perfect discrimination.
A valid determination of the prediction tool’s discrimination would require the application of the prediction tool in populations other than the population used in development of the prediction tool. Thus, the discrimination should be ideally tested in an independent cohort; however, in the absence of an external cohort, prediction tools are usually subjected to internal validation. Internal validation can be done by bootstrapping (the development dataset is used to simulate prediction tool testing under novel conditions) [23–28], split-sample validation, and/or cross-validation (such as 5-fold or 10-fold cross-validation or leave-one-out validation) .
While discrimination quantifies the ability of a prediction tool to distinguish between patients, calibration quantifies the accuracy of a prediction for an individual patient. A tool is well calibrated if close to x% of patients given an x% risk have the event. Calibration plots graphically illustrate the relationship between predicted and observed rates of the outcome of interest. Ideally, a well calibrated prediction tool exhibits a 1:1 relationship between predicted and observed rates, resulting in 45 degree slope (e.g., Figure 2B). Such plots can be calculated for the data set used to create the model (in which case poor calibration is rare) or for external datasets.[24, 26–29].
Differences in disease and population characteristics may undermine the discrimination and calibration of predictive tools when applied to a different population. Specific criteria used in defining the sample used to develop a prediction tool may not allow the use of tools for patients with different characteristics or who have been exposed to different treatment modalities. For example, a prediction tool that is specific to screening population may perform differently when applied to a referral population . Similarly, prediction tools that were developed using high-volume single center databases may not be applicable to community practice. Therefore, it is important that clinicians judge whether a specific prediction tool is indeed generalizable and applicable to the population they intend to apply it to [24, 26–28]. Models that do adjust for important factors such as grade and stage are more likely to be generalizable. A model could be useful if it adjusted for the major characteristics.
A prediction tool’s ability to predict a specific outcome may be affected by population characteristics that change over time. In general, more contemporary PCa patients are diagnosed with more favorable stage and grade. Therefore, prediction tools require periodic reappraisals to assess the effect of change of stage and grade on the prediction tool. One may find that predictions devised on historic cohorts are less accurate for contemporary patients. Some prediction tools, however, have shown stable performance characteristics such as discrimination and calibration . External validation in contemporary cohorts is necessary to ensure continued validity.
The level of complexity of a predictive tool represents an important practical consideration. One has to distinguish between complex algorithms and models with complex inputs required. Complex algorithms have fewer problems because in the future the calculations will be done by computers. However, models with complex inputs are clearly impractical in busy clinical practice. This may, however, become a moot point with the advent of “smart” medical records. On the other hand, the use of variables that are not routinely available, such as novel molecular markers, is impractical.
Because of the protracted course of prostate cancer, prediction tools may need to take competing causes of mortality into consideration. Consideration of competing risks is a feature of several prediction tools for patients with bladder cancer.[31–34] Competing risk modeling is able to predict cancer control rates after accounting for the effects of competing risks. Such modeling is important if risk of non–cancer related death is high, or if predictors of death overlap with predictors of the event of interest. If risk of death is relatively low, and predictors of death are separate from predictors of the event of interest, then adjustment for competing risks will have little effect on prediction. For PCa, it seems that there is need to better understand PCa-related risks in the context of other-cause mortality. Predictions that account for competing risks are important to clinicians, as well as to patients, especially when there is danger of over-treatment or sub-optimal treatment. Indeed, since the morbidity and mortality of PCa treatment are not trivial, clinicians must be able to better risk-stratify PCa patients to ensure that treatment is directed toward those who are most likely to benefit from it .
A patient’s probability of a future event changes over time; for example, a patient’s risk of disease recurrence within 5 years after radical prostatectomy is higher the day after surgery than if the patient has an undetectable PSA at 48 months. In PCa, because patients’ prognosis is expected to improve with increasing disease-free interval, absence of adjustment for this variable results in an excessively somber estimate of cancer control over time. The updated versions of the pre- and post-operative Kattan nomograms for prediction of biochemical recurrence provide the opportunity to adjust for the effect of disease-free interval from surgery [36, 37]. As expected, the predicted risk of disease recurrence decreases with increasing disease-free interval.
When judging a new tool, one should compare it with established prediction tools, in order to determine whether the new prediction tool offers advantages over the available alternatives [27, 28, 38–44]. The most direct and unbiased criteria for comparison are the objective attributes of discrimination and calibration. Subsequently, complexity, generalizability, and other attributes may also be compared. This sort of head-to-head comparison is a better approach than judging the concordance index or AUC in isolation or against a possibly arbitrary threshold.
In a head-to-head comparison, each prediction tool being compared is applied to a common external dataset. The prediction tools are applied to each individual observation to derive the probability of the outcome of interest. The predictions are then compared against observed rates of the outcome of interest, and discrimination is calculated using the receiver operating characteristics curve. A common mistake consists of refitting a prediction tool on the same dataset on which it was constructed and describing this as external validation.
While discrimination and calibration are important, methods that incorporate clinical consequences are crucial to the evaluation of prediction tools. This type of analysis allows insight into the consequences of using a test, prediction tool, or marker in the clinic. Several methods are available. Decision curve analysis is a method that combines simplicity with efficient computations [45–48]. In brief, the method is based on the principle that the relative harms of false positives (e.g. unnecessary biopsy) and false negatives (e.g. missed cancer) can be expressed in terms of a probability threshold. For example, if a man would opt for biopsy if he was told that his risk of PCa was 20% or more, but not if his risk was less than 20%, it can be shown that he considers that harms associated with a delayed diagnosis of cancer to be four times greater than the harms associated with an unnecessary biopsy (i.e. the ratio of harms is the odds at the probability threshold). This threshold probability can therefore be used both to determine whether an individual patient’s test result should be defined as positive or negative and to weight the clinical consequences of true and false. An example of a decision-curve is shown in Figure 1. The decision analytic evaluation should be performed during later stages of research and before clinical implementation of the tool.
Various distinct statistical methodologies have broadly been described as “nomograms.” Although widely used as a synonym for “prediction model”, the strict definition of a nomogram is a graphical calculating device that can be based on any type of function [12, 23]. The nomogram usually incorporates at least two variables, which can be either continuous or categorical. The effect of the variables on the outcome of interest is represented by axes, and risk points are attributed according to the prognostic significance of the variable of interest. For example, in the nomogram in Figure 2 , each PSA value is assigned a unique point value that represents its prognostic significance. The ‘Total Points’ axis is used to estimate the combined effect of all predictors on the probability of biochemical recurrence. One advantage of this format is that it allows the input of continuously coded variables as well as categorically coded variables, without limit on the number of variables. This format distinguishes nomograms from look-up tables or decision trees, where continuously coded variables cannot be processed and where data availability limits the degree of stratification because of the need to avoid empty cells or dead-end branches.
Nomograms are designed to extract the maximum amount of useful information from the data. For example, the primary and secondary Gleason grades are used as independent variables, rather than the Gleason sum alone. This additional information is useful because several combinations of primary/secondary Gleason grades can result in the same Gleason sum (e.g., 3+4=7 vs 4+3=7) , despite a different impact on prognosis.
The available PCa nomograms have been adapted for use on personal digital assistants and personal computers to facilitate their integration into daily clinical practice and research. Many of the nomograms can be found either on the Memorial-Sloan-Kettering Cancer Center (www.nomograms.org) or the University of Montreal (www.nomogram.org) websites.
The above discussion is meant to provide guidelines for analyzing and using predictive tools. Next, we provide an overview of predictive tools available for PCa. We describe predictor variables, the outcome of interest, the number of patients utilized to develop the tools, tool-specific features, predictive discrimination estimates, and whether internal and/or external validation has been performed. We focus on nomograms because we believe that they represent one of the most accurate tools to date.
Table 1 shows tools for prediction of PCa on initial biopsy. Eastham et al. developed the first such tools, yielding a c-index of 0.75 . Despite good discrimination, this tool suffers from limited generalizability. The tool was limited to men with suspicious digital rectal examination and serum PSA <4.0 ng/ml. In addition, the tool predicts only the results of sextant biopsies, further limiting its applicability. Chun et al. developed and validated tools in a population subjected to extended biopsy sampling . External validation in three cohorts totaling 2900 men demonstrated 73 to 76% discrimination. Nam et al. developed a prediction tool that predicts an individual’s risk for PCa in a cohort of 3,108 men who underwent a prostate biopsy for the first time. The dataset included a subset of 408 volunteers with normal PSA levels. The tool comprises factors that can be easily determined at the time of screening such as age, ethnicity, family history of PCa, the presence of urinary symptoms, PSA, percent free PSA, and digital rectal examination. The prediction tool’s discrimination was 74% and 77% for high-grade cancer (Gleason score ≥7). This was significantly greater than the AUC that considered using the conventional screening method of PSA and DRE only (62% for any cancer and 69% for high-grade cancer).
The PCPT investigators developed a risk calculator from a large randomized trial with an AUC of 0.70 for the ROC curve in the original cohort and 0.65 in the independent validation cohort.[54, 55] This study is a potentially helpful additions to the growing number of prostate cancer nomograms
Table 2 shows tools for prediction of PCa on repeat biopsy, saturation and mixed indication (initial and repeat biopsy). A tool described by Lopez-Corona et al. was developed and internally validated (70% discrimination) in 343 men  and externally validated (71% discrimination) in 230 men . Chun et al. developed the most contemporary repeat biopsy tool (n=1162) based on 10 or more biopsy cores . In three external validation cohorts (n=582, n=961, and n=195) the nomogram yielded 73–76% discrimination.
Several multivariable statistical prediction tools have been proposed to estimate pathologic stage at radical prostatectomy (RP) with the intent of facilitating treatment planning (Table 3). Of these, the “Partin tables” represent the most widely used tool. These look-up tables predict pathologic stage at RP . After their introduction in 1993, the Partin tables were validated in 1997 and updated in 2001 and 2007 [19, 59–61]. Although the Partin tables represented a milestone in pretreatment PCa staging, they have limitations. For example, the probability of extracapsular extension (ECE) cannot be predicted in a side-specific fashion. To circumvent this limitation, Graefen et al. devised a regression tree capable of predicting ECE in a side-specific manner . This prediction tool allows the identification of candidates for non–nerve sparing RP versus unilateral versus bilateral nerve-sparing RP. External validation of this prediction tool yielded 70% discrimination 
Ohori et al. developed a tool (n=763) to predict side-specific ECE (range of c-index: 79–81%) . Validation of another side-specific prediction tool in 1118 European patients yielded 84% discrimination . Compared to the Partin tables, the tool approach is more accurate and provides side-specific predictions. Moreover, the tools predict ECE independently of seminal vesicle invasion (SVI) and lymph node invasion (LNI) [18, 19, 61].
Koh et al.  and Gallina et al  used data from 763 patients to devise tools to predict the probability of SVI (range of c-index: 78–88%). Cagiannos et al. developed an LNI prediction tool (n=5,510) yielding 76% discrimination, compared to 74% for the Partin tables tested in the same cohort . Briganti et al. developed an LNI prediction tool from 602 patients who underwent extended pelvic lymphadenectomy (76% discrimination) . In addition, the same investigators developed another tool (n=565) that allows identification of patients who are at a negligible risk of LNI outside the obturator fossa (80% discrimination) and therefore may not require extended pelvic lymphadenectomy . The combination of these tools allows accurate identification of the need for and extent of pelvic lymphadenectomy.
While prediction of adverse pathologic features is important for management of PCa patients, a proportion of patients harbor clinically insignificant or indolent PCa, which cannot be predicted with these tools . To address this void, Kattan et al. developed three prediction tools (range of c-index: 64–79%) that predict the probability of indolent PCa  based on the definition of Epstein et al. (i.e., organ-confined cancer, 0.5 cm3 or less in volume, and no poorly differentiated elements) . Steyerberg et al. externally validated these tools in a screening cohort (range of c-index: 61–76%) . These models can reasonably predict the likelihood that a patient has a small, well to moderately differentiated PCa that is confined to the prostate. While pathologic outcome may not define the biological potential of an individual tumor, it provides valuable information to a patient who is considering conservative management of PCa.
These tools allow prediction of clinically insignificant PCa with reasonable discrimination, which may help in deciding between definitive therapy and active surveillance. Conservative management may be appropriate in patients with a high probability of indolent cancer, eg exceeding 60%. In those with a low probability, eg less than 30%, potentially curative management may possibly be advised. Note that these cutoffs for the probability of indolent cancer are speculative and further research is required to give more insight into the pros and cons of prostatectomy in men with a substantial probability of indolent cancer. Other factors may dominate the decision making process in individuals.
Kattan et al. developed the first pre-treatment prediction tool (n=983), which predicts the 5-year biochemical recurrence for patients treated with RP . External validation yielded accuracies of ~75% (range: 65%–83%) [72–74]. However, the 5-year endpoint is insufficient to predict the likelihood of cure after RP, as patients are still at risk of disease recurrence beyond 5 years [75–77]. Therefore, Stephenson et al. recently updated the pre-operative prediction tool by predicting the 10-year probability of biochemical recurrence after RP (Figure 2A; 77% discrimination) . The prediction tool exhibited good calibration across the spectrum of predictions in internal validation but exhibited some optimism in external validation (Figure 2B). An added feature of the tool is the ability to estimate the probability of recurrence at any point in time from 1 to 10 years after RP, accounting for disease-free interval.
Kattan et al. also developed a post-operative tool for prediction of 5-year biochemical recurrence using data from 996 men treated with RP for clinically localized PCa by a single surgeon (73% discrimination) . External validation yielded accuracies of 80% (range: 77%–82%) in an international cohort  and 83% in an African-American cohort . Stephenson et al. updated the post-operative tool by including contemporary patients and extending predictions up to 10 years after RP while accounting for disease-free interval (Figure 3) . External validation yielded a discrimination of 78% to 81% . Suardi et al. developed the furthest-reaching prediction tool, which provides the probability of biochemical recurrence up to twenty years after RP. Their prediction tool also accounts for disease-free interval. The prediction tool’s discrimination (77–83%) was confirmed in two external validation cohorts .
An important aspect of pre-treatment counseling is pre-treatment estimates of the outcomes if treated with three-dimensional conformal external beam radiotherapy or brachytherapy. Table 5 shows prediction tools that predict biochemical recurrence after radiation therapy, all of them using the original American Society for Therapeutic Radiology and Oncology (ASTRO) definition of biochemical recurrence as 3 consecutive increases in PSA . Kattan et al. developed a pre-treatment prediction tool to predict the 5-year biochemical recurrence-free probability after three-dimensional conformal external beam radiotherapy (n=1042; Figure 4) . External validation within a cohort of 912 men yielded a discrimination of 76% [82, 83].
The same authors developed a prediction tool that predicts 5-year biochemical recurrence-free survival after 125I brachytherapy without adjuvant hormonal therapy (n=920, Figure 5) . Two separate external validations resulted in discriminations of 61% (n=1827) and 64% (n=765). Minor differences in the population characteristics account for the different discrimination in the different validation populations.
To address metastatic progression after definitive therapy, Kattan et al. developed a prediction tool quantifying the probability of metastatic progression within five years following external beam radiotherapy (n=1,677) . The prediction can be made immediately after definitive therapy. Its external validation demonstrated a discrimination of 81% (n=1,626).
Dotan et al. developed a tool to predict the probability of metastatic progression, defined as a positive bone scan, in 239 men with a rising PSA after RP . The tool relies on detailed serum PSA characteristics including kinetics and therefore requires multiple post-recurrence PSA values. The prediction tool discrimination is 93%.
Slovin et al. devised a similar tool predicting the time to radiographically detectable metastases in patients with biochemical recurrence (n=148) after either RP or external beam radiotherapy. Similar to the prediction tool from Dotan et al., the tool from Slovin et al. requires the consideration of PSA kinetics in the form of PSA doubling time and is limited to patients whose doubling time is less than 12 months . The predictive discrimination of this prediction tool was 69%, but it was neither internally nor externally validated.
Armstrong et al. developed a tool using baseline clinical variables to predict death among men with metastatic hormone-refractory prostate cancer (n=1,006) using data from the TAX327 randomized clinical trial (three week or weekly docetaxel or mitoxantrone, each with prednisone). Ten independent prognostic factors other than treatment group were included: presence of liver metastases, number of metastatic sites, clinically significant pain, Karnofsky performance status, type of progression (measurable disease progression and bone scan progression), pretreatment PSA doubling time, PSA, tumor grade, alkaline phosphatase, and hemoglobin. The predictive discrimination of this prediction tool was 69%. External prospective validation may support the wider use of this prognostic baseline model for men with HRPC treated with chemotherapy.
Four tools have been devised for prediction of survival in PCa patients. One predicts cause-specific survival in PCa patients exposed to hormonal therapy, regardless of the time of hormonal therapy initiation. The remaining three tools predict the probability of all-cause survival in patients with androgen-insensitive PCa (AIPC).
Porter et al. developed a tool for prediction of cause-specific survival in patients exposed to hormonal therapy after RP failure (n=114) . The internally validated discrimination of the tool was only 66%. Among the three prediction tools for AIPC patients, those of Smaletz et al  and of Halabi et al  were developed and externally validated in heavily pretreated patients, who had been exposed to one to several experimental agents. In external validation, the accuracy of the prediction tool of Smaletz et al. was 67% and that of Halabi et al. was 67% and 68%.
Svatek et al. devised a contemporary prediction tool using a population with a median survival of 52 months who had not received experimental therapies. This prediction tool relies predominantly on PSA doubling time and PSA level at hormone therapy initiation. Internal validation of this prediction tool yielded a discrimination of 81%. The contemporaneity and homogeneity of the patient population make this prediction tool very attractive when survival needs to be assessed in patients with AIPC.
Life expectancy is a crucial consideration in treatment decisions, especially in patients who are candidates for definitive therapy. Ten years is generally accepted as the minimal life expectancy pre-requisite for curative local treatment in men with PCa. Unfortunately, life tables are not reliable for prediction of life expectancy in definitive therapy candidates, demonstrating only 60% discrimination . Moreover, clinicians are also poor raters of life expectancy, with 19 clinicians yielding only 69% discrimination. To circumvent this problem, Tewari et al. (n=3159) , Cowen et al. (n=506) , and Albertsen et al. (n=451) [95, 96] developed prediction tools predicting life expectancy in men with PCa. The discrimination of these prediction tools ranged from 69% to 73%. Unfortunately none of the three prediction tools is capable of discerning between PCa-specific and all-cause mortality. To address these considerations, Walz et al. devised a tool for prediction of life expectancy in excess of 10 years in candidates for RP and external beam radiotherapy (n=9131) . The prediction tool focused on patients without evidence of cancer relapse after definitive therapy. The predictors consist only of age and co-morbidities. Internal validation of this tool yielded 84% discrimination vs. 81% for the prediction tool of Tewari et al. Based on discrimination, simplicity, and generalizability, the prediction tool of Walz et al. appears to provide the best estimates of life expectancy in excess of 10 years.
The discrimination of current predictive tools is imperfect. To date, discrimination has not been significantly improved by the addition of other potentially informative clinical and pathologic features [98, 99]. Discrimination may, however, potentially be improved by incorporation of novel biomarkers and/or imaging data that are associated with the biologic behavior of PCa.
Over the past two decades, the molecular dissection of cancer has increased our understanding of the pathways that are altered in neoplastic cells. Protein expression profiling of PCa offers an alternative means to distinguish aggressive tumor biology and may improve the accuracy of outcome prediction. In addition, such markers may serve as prognosticators and/or therapeutic targets that facilitate the emergence of new therapeutic approaches.
However, despite numerous reports of promising new biomarkers in the urological literature, to date only a few studies have demonstrated an important improvement in predictive discrimination when biomarkers were added to established predictors in prediction tools (example: Figure 6) [43, 97, 100–105].
Kattan et al.  developed and internally validated a prognostic prediction tool that incorporates pre-operative plasma levels of transforming growth factor-β1 and interleukin-6 soluble receptor in the standard pre-operative Kattan nomogram (which uses pre-operative PSA, biopsy Gleason grade, and clinical stage to predict the probability of biochemical recurrence-free survival at 5 years after RP) . Addition of these biomarkers improved the predictive accuracy by a statistically and prognostically substantial margin relative to clinical variables alone (increase in accuracy from 75% to 83%). This prognostic tool was externally validated in an independent cohort comprising 423 consecutive men who underwent radical prostatectomy . The validation prediction tool provided a discrimination of 87.9% versus 71.1% for the tool including clinical variables only (16.8% difference; P < .001). Moreover, the prediction tool was better calibrated than the tool including clinical variables only. These prediction tools further refine our ability to identify patients at a high risk of biochemical recurrence after radical prostatectomy.
Non-invasive diagnostic imaging, especially magnetic resonance imaging (MRI) and MR spectroscopic imaging (MRSI), has improved in recent years and is gaining widespread acceptance for aiding PCa diagnosis, tumor localization, staging, assessment of tumor aggressiveness, and treatment planning. Investigators have used nomograms and neural network prediction tooling to predict organ-confined PCa , clinically significant disease , and biochemical recurrence after RP .
The paucity of randomized trials in PCa makes diagnostic and treatment decisions complex. Prediction tools can provide predictions that are evidence-based and at the same time individualized. Such predictions have been repeatedly shown to be more accurate than those of clinicians, regardless of their level of expertise [7, 10]. Prediction tools such as nomograms have been embraced by the urologic community. The nomogram format has also been adopted in several other disciplines of oncology, including breast, colon, bladder, gastric, and lung cancers.
PCa patients need to be involved in decisions regarding management of their disease. They should know what their options are and what the consequences can be. At the core of any patient involvement would be accurate prediction of consequences and, essentially, a spreadsheet of these predictions tailored to the individual. This spreadsheet would help make consent for any medical decision truly informed. Providing this information could reduce the likelihood of regret of treatment choice, particularly when complications arise.
Continuous multivariable prediction tools such as nomograms are a highly appealing means of calculating accurate predictions with or without the use of a computer. Nomograms currently represent one of the most accurate and discriminating tools for predicting outcomes in patients with PCa [25, 39–41, 109–111]. When faced with the difficult decision of choosing among the treatment options for each clinical stage of PCa, patients benefit from the accurate estimates of outcomes provided by predictive tools. Equipped with these predictions, patients are more likely to be confident in their treatment decisions and less likely to experience regret in the future. However, we emphasize that the predictions must be interpreted as such; they do not make treatment recommendations or act as a surrogate for physician-patient interactions, nor do they provide definitive information on symptomatic disease progression or complications associated with treatments.
The fundamental issue of predictive tools such as nomograms pertains to their utility. Indeed, very limited data exist with respect to the impact of nomograms on medical decision-making. Although studies have shown that decision aids improve patient knowledge and affect medical decision-making behavior , the role of tools has yet to be studied sufficiently. There are no prospective randomized studies that clearly demonstrate that the use of prediction tools improve patient care or reduce patient anxiety, decisional conflict, or regret. A clinical trial would be very valuable for establishing the effects of prediction tools; however, whereas informing patients with predictions regarding the impact of a medical procedure seems ethical, withholding accurate outcome predictions from patients in a control group in a randomized trial does not. Such a trial would likely be associated with significant power problems, on the grounds that a nomogram would be worth using even for a small improvement in outcome. Presently, patients are using very limited quantitative information when making their decisions, and direct outcome predictions are the simplest factors for them to consider. Nevertheless, prospective evaluation of prediction tools and decision analytic methods would be beneficial.
Besides being required for improved medical decision-making, accurate risk estimates are also required for evaluation of novel markers and for design of clinical trials, for example to enable the selection of homogeneous groups of patients as study participants. Prediction tools have the potential of improving the ability of Phase II trials to discriminate between ineffective and potentially effective therapies. Prediction tools may help enroll populations that are homogeneous with respect to risk. In addition, they would allow comparison with other studies and historical controls based on risk. For Phase III trials, prediction tools can help to ensure that eligible patients have sufficiently high levels of risk, thereby increasing event rates and reducing sample size requirements. Risk prediction tools define high-risk patients more accurately than do risk-grouping strategies [39, 109]. Use of risk predictions for individual patients, therefore, decreases the proportion of low-risk patients enrolled, avoiding unethical inclusion, as well as increasing statistical power. Finally, future designs of Phase III trials should include prediction tools to increase the clinical utility of their findings. We therefore recommend the wider adoption of risk prediction tools and decision analysis in the design, analysis, and implementation of clinical trials.
Areas of future research include the integration of genomic/molecular markers and development of prediction tools that predict additional clinically significant endpoints such as clinical recurrence, cause-specific death, long-term survival, and outcomes related to morbidity and quality of life. Models that use updated post-treatment information based on biomarkers such as PSA will and do play a very important role in decisions about new treatment when recurrence is suspected.
Better modeling may allow for adjustment of the trend of decreased accuracy (discrimination ~ 60–70%) for “longer” predicted endpoints such as metastasis and survival compared to those for diagnosis, stage of PCA or disease recurrence (discrimination usually better than 70%). Moreover, with the advent of novel therapies such as laparoscopic/robotic prostatectomy, high-intensity focused ultrasound, and focal ablative therapy, tools specific to these procedures are needed.
Patients with PCa are heterogeneous with respect to disease aggressiveness, age, comorbidities, and personal preferences. Therefore, to obtain optimal cancer control with minimal side effects, treatment should be tailored to each individual patient and the characteristics of his tumor. In order to make an informed decision, many patients want to know their likely outcomes, and clinicians need to provide accurate estimates of these outcomes. Despite limitations, predictive tools can provide individualized, evidence-based estimates for a number of PCa endpoints, thereby helping in the complex decision-making process. Moreover, predictive tools provide an objective justification for clinical decisions, which is increasingly demanded by patients, administrators, peers, and third-party payers.
1. The Sidney Kimmel Center for Prostate and Urologic Cancers.
2. T32 grant from the National Institute of Health (T32CA082088).
We thank Janet Novak, PhD, for substantive editing of the manuscript. This work was paid for by MSKCC.
Conflict of Interest Statement
MWK and PTS are co-inventors of several patents and patent applications (20070111269 Method to predict positive repeat prostate biopsy; 20050282199 Method to predict prostate cancer; 20030235816 Method to determine outcome for patients with prostatic disease).
SFS is co-inventor of two patent applications (20030235816 Method to determine outcome for patients with prostatic disease; 20030054419 Method to determine prognosis after therapy for prostate cancer).