Jan Verbeek

DISCUSSION AND FUTURE RECOMMENDATIONS

CONCLUSIVE

In this thesis, five research questions were answered related to two major topics. The first topic addressed the research results of prostate cancer screening and active surveillance. It explored the translation of the relative risk reduction of the screening into the real-life scenario of the absolute risk reduction by risk-stratification. The second topic evaluated the subject of PCa screening from a personalized medicine perspective. More specifically we asked: can we reduce unnecessary testing using multivariable prostate cancer prediction models, and can risk-stratified models be of value in clinical decision-making?

Classic evidence-based medicine: Outcomes of prostate cancer screening and active surveillance

Question 1 – What are the risks of a diagnosis of a clinically significant PCa, metastatic disease and/or PCa death after a false negative screening test or biopsy result in a purely PSA-based screening setting compared to applying an additional test procedure or risk stratification tool? (Chapter 2)

There is strong evidence that PSA screening by applying systematic prostate biopsies for test positivity criterion ≥3.0 ng/mL is beneficial in achieving a 20 % relative risk reduction of PCa death. A logical question is whether a larger benefit is possible by addressing the false negative rates for men with PSA <3.0 ng/mL and negative biopsy. We discovered that these false negative rates are extremely low and PSA (repeat) screening (including sextant biopsies) detects almost every PCa case. Therefore, additional tools like biomarkers and mpMRI are unlikely to detect more clinically relevant PCa. Nonetheless, these additional tools may be useful for risk-stratification and have added value in reducing the high rates of unnecessary biopsy (no cancer) and considerable overdiagnosis of low-risk PCa. Question 2 – What is the compliance over time when offering an AS protocol to men with low-risk PCa and how can risk stratification at the start of AS optimize adherence? (Chapter 3) AS is based on the concept that low-risk PCa is unlikely to harm or decrease life expectancy. Management of slow growing PCa with AS is a better choice than immediate active treatment with surgery or radiation, including complications and side effects. The strategy is substantiated by studies that show that men with low-risk prostate cancer who have been on AS for 10 - 15 years after diagnosis have extremely low rates of disease spread or death from prostate cancer. In addition to the definition of low-risk and/or indolent PCa, the eligibility and inclusion criteria for AS, the AS protocol itself, and adherence to the protocol is key determinant of progression of disease. The Movember Foundation's Global Action Plan Prostate Cancer Active Surveillance Initiative (GAPP initiative) provided results from 25 AS cohorts worldwide. Adherence to the annual schedule of repeat biopsy reduced with time: we estimated a pooled one-third nonadherence after 7 years. For making evidence-based AS selection guidelines, the importance of risk stratification with age, tumor stage, PSA level, and diagnosed biopsy score was stressed to optimize future AS protocols, mpMRI surveillance was also considered. Overall, constructing a more personalized risk-based approach to biopsy scheduling on AS was recommended. Towards personalized medicine: How can multivariable prostate cancer prediction models reduce unnecessary testing and support clinical decision-making? Question 3 – How do prediction models work and how should the prediction outcomes be interpreted in terms of clinical utility? (Chapter 4) Classical evidence-based medicine methods analyze the benefits and harms of tests and treatments in terms of relative risk reduction and/or elevation. The concept of numbers needed to screen, biopsy, diagnose, treat, and harm is used as a better way of communicating and clinical decision-making. A number needed to screen of, say, 570 means that the death of 1 in 570 men will be prevented through PCa, and the remainder will not. Unfortunately, we do not yet know who the ‘lucky’ one will be. Prediction models address the unique situation for the individual man. These models are designed and include patient characteristics, test results, biomarkers and imaging to accurately predict the occurrence of defined endpoints. The clinical impact of newly developed risk prediction models is currently assessed with decision curves. In the ‘statistics in urology review’ presented in Chapter 4, we made a plea for investigators with regards to reporting and correctly interpreting decision curve analysis. We present a statistical method to evaluate whether the model is useful in clinical decisions, and whether extended models with, for example, innovative biomarkers to predict high-grade PCa, will lead to better decisions. Question 4 – Can prediction model predicting biopsy outcome be improved by incorporating novel biomarkers and a more refined PCa pathological grading system, and hence decrease the number of unnecessary prostate biopsies and overdiagnosis of potentially indolent disease? (Chapters 5 and 6) Strong evidence that primary-based PSA testing works to reduce PCa mortality has been presented. However, side effects are unavoidable because a raised test result does not necessarily mean a PCa diagnosis, so further examination is needed. The PSA test has a high false positive rate (the predictive value positive of prostate biopsy is low), indicating a situation of too many unnecessary biopsies. Prediction models are adopted to refine diagnosis (differentiation of potentially lethal PCa from relatively indolent cancer) and reduce the number of unnecessary biopsies. The models incorporate patient characteristics, biomarkers and technology to optimize the balance between the likelihood of benefit and the risk of harm for individuals. We studied the prediction model underlying the No. 3 Rotterdam European Randomized Study of Screening for Prostate Cancer risk calculator for initial prostate biopsy and compared it with an updated risk calculator augmented with the contemporary Gleason grading and cribriform growth pathological biopsy classification. This Rotterdam calculator will lead to a 34% reduction in unnecessary biopsy, while only 2% of high-risk PCa will be undiagnosed. We then searched for a novel biomarker to detect PCa to be included as a risk calculator able to deliver optimal net benefit outcome, the 4-kallikrein panel (4K-score). The prediction performance was studied with discrimination and calibration plots, and decision curve analysis was used to evaluate the reduction of unnecessary biopsy and indolent PCa. Compared to PSA testing, the Rotterdam Prostate Cancer Risk Calculator and the 4K-score equally reduce the number of biopsies by approximately two-thirds. Prostate cancer prediction can be slightly improved by combining the Rotterdam calculator with the 4K-score. Question 5 – To what extent can prediction models support triage at primary care practice regarding who receives screening and diagnostic examination, thereby reducing unnecessary testing and overdiagnosis? (Chapter 7) Current consensus-based guidelines advise not to refer men for biopsy if aged >75 as they have a low life expectancy. When making these decisions, it is relevant to know that 50% of Dutch 75-year-old men are expected to live more than 11.5 years, but also that approximately 25% of them will live for more than 15 years, while only 25% will live less than 6 years. We therefore recommend adding to clinical usefulness of the tool’s information on risk of cancer and life expectancy.

A tool for shared-decision-making on referral for biopsy in the primary care setting was developed. Data and estimates were collected from the Dutch arm of the ERSPC trial, treatment trials, cancer registries, and national mortality statistics. A negative impact on life expectancy and treatment benefit was found with higher age and more comorbidity. The proposed multivariable and multidimensional prediction tool with information on life expectancy, risk of aggressive PCa, and potential benefit of prostate cancer treatment comorbidity needs further validation. We are confident that it can provide general practitioners and their patients with more accurate information regarding whether or not the patients should be referred for prostate biopsy.

FUTURE PERSPECTIVES

Developments in the early detection of PCa and prostate cancer care go fast. Classic evidence-based medicine applies to the average population and cannot simply be translated into person-level predictions, which is the objective of personalized medicine. How the future will unfold can only be presented in general terms. Table 2 provides points:

o Classic evidence-based medicine applies to the average population and cannot simply be translated into person-level predictions
o The current state of evidence on prostate cancer screening is still regarded insufficient to start large-scale screening programs
o Updates with novel biomarkers and imaging techniques could favor the discussion to start with screening programs and should be weighed with other significant improvements in terms of operating techniques, systemic treatments, localized radiology treatment as they also influence PCa mortality
o Active surveillance is a safe option for men with low-risk PCa, however the definition of low-risk PCa should be defined in absolute risks of clinically significant PCa
o The shift towards personalized medicine with prediction models provides more patient-specific intervention estimates which support individualized clinical decision-making, instead of using a relative risk of intervention.
o Decision curve analysis is introduced as a novel method for evaluating the clinical usefulness of prediction models to aid patients with their decisions, however, there is still a way to go.

Prostate cancer screening
The current state of affairs and body of evidence on prostate cancer screening is still regarded insufficient to start large-scale screening programs [51]. Still, prediction models that include biomarkers and imaging findings are promising and increase the net benefit of prostate cancer screening, mainly by reducing the associated harms. However, it has to be kept in mind that most evidence is still based on small population and patient cohorts, and often on retrospective studies. In recent years, many companies have developed novel biomarkers such as the 4Kscore (OPKH Health Inc), PHI (Beckman Coulter), SelectMDx (MDxHealth Inc), and ExoDX (Bio-Techne). In addition, there have been great improvements in histopathological features and imaging techniques like TRUS, PSMA-PET/CT and mpMRI techniques (Philips, Siemens and General Electrics). Before a proper screening program can be implemented, hard evidence on the predictability of these biomarkers, histopathological features and imaging techniques needs to be gathered based on new, large screening trials powered on intermediate endpoints or cohort studies opting for predicting modeling and personalized medicine. Moreover, these predictors should be compared with each other to improve the total clinical utility. A new Finnish screening study is currently under way, conducted with 4Kscore and mpMRI [52], but before the results can be interpreted more participants and longer follow-up are required [53]. It should be realized that our understanding of the benefits of prostate screening comes from 25-year-old data, and updates are strongly needed, as significant improvements in terms of operating techniques, systemic treatments, localized radiology treatment also influence PCa mortality. As times moves on, even the data resulting from a new screening trial will become outdated, thus instead of continually initiating a new trial, better registration of large cohorts is a good alternative. Cost-effectiveness considerations are also important, not only when making decisions regarding launching a population-wide screening program, but also in individual testing. Studies comprising cost-effectiveness analysis are required when improvements in clinical decision-making are expected.

Active surveillance
AS is a safe option for men with low-risk PCa, however the definition of low-risk PCa should be defined in absolute risks of clinically significant PCa, indolent cancer or benign hyperplasia instead in relative risk reductions to provide appropriate care for each individual. In the near future, clinical and histopathological features, biomarkers, and imaging techniques should be used in a complementary manner in multivariable prediction models. These models may achieve optimal risk stratification and maximize the effects, e.g. the avoidance of detection low-risk PCa.

Prediction models
We now live in an era where vast amounts of data can be stored and quickly processed, with new advances in developing and validating accurate prediction models. What was first a simple linear correlation can now be a higher order transformation with restricted cubic spline with multiple knots, or even more flexible exploitation of correlation with machine learning methods and artificial intelligence. However, we should be aware of the problems related to overfitting which occurs when a prediction model is too complex to be developed in a specific sample of limited size. Moreover, transportability may be limited when a model is implemented a clinical setting other than the setting for which it was developed [10]. Overfitting can be prevented by using large numbers in the development of the prediction model, but also with sensible modeling [54]. Sensible modeling means applying external knowledge and using a model only for a specific calculation. Also, as overfitting automatically occurs shrinking the model should be applied. This is done by estimating the overfitting factor with cross-validation or bootstrapping techniques. Finally, external validation is used to see the heterogeneity between different settings, if the heterogeneity is low the model is well transportable. Instead of statistically measuring how well a prediction model works, the clinical outcome can be simulated before implementation, the first step in this process is possible by using the Net Benefit approach [27]. The shift towards personalized medicine provides more patient-specific intervention estimates which support individualized clinical decision-making, instead of using a relative risk of intervention [55]. Therefore, healthcare with predictions models will benefit patients and provide an overall higher clinical utility.

Aiding patients with their screening decisions
In modern medical practice, decisions and interventions can benefit from the results of comparative group research. Using group comparison, it is possible to quantify the relative benefit for the individual. For example, the number needed to screen indicates how many people need to comply with screening in order to avoid one PCa cancer death. As a doctor aiding his or her patients with the decision to screen or not to screen, more information is needed. Decision curve analysis was introduced as a novel method for evaluating the clinical usefulness of prediction models, and adopted gradually by the urological community. However, there is still a way to go. The initial authors developed the tutorial because investigators indicated that decision curve analysis is difficult to understand, most probably as “the two axes of the decision plot —threshold probability and net benefit— are concepts that are novel to many” [56]. They argued that “many of the difficulties in interpreting decision curves can be solved by relabeling the y-axis as “benefit” and the x-axis as “preference”, i.e. a new x-axis ranging from “I’m worried about disease” towards “I’m worried about biopsy”. Still uncertainty about the future is considered to be the most common source of stress in humans apart from traumatic stress and can be difficult to put in the perspective of risk prediction and risk management for most people. Time will tell how instruments for shared-decision-making will evolve.