Prostate cancer is a growing burden on health care systems worldwide [1, 2]. With rising disease awareness, an increasing proportion of men are presenting with non-metastatic disease [3, 4]. There is an urgent need to improve the prognostic precision for men with non-metastatic disease since management options are becoming more diversified, e.g. the increasing use of active surveillance for low-risk disease, and, conversely, due to the recognition that more intensive, combined treatment is needed in high-risk disease [5, 6]. Current risk stratification models were primarily developed to predict therapy failure and not the risk of prostate cancer death. Moreover, they are almost exclusively based on surgically and radiotherapy-treated men and do not include men who are managed conservatively [7, 8]. Nevertheless, the simple clinico-pathological variables that go into these models make them easy to use, and they are commonly the first triaging step recommended by many national and international guidelines for clinical decision-making [9–12].

To address this, we recently remodelled the components (histological grade, clinical stage and prostate-specific antigen (PSA) at diagnosis) that comprise the currently used risk classification systems [13]. In a new five-strata model, we also incorporated the new histological grade grouping recently recommended by the International Society of Urological Pathology (ISUP), which has been shown to be a better predictor of disease recurrence and progression than the Gleason sum alone [14, 15]. In a cohort of nearly 12,000 UK men, we found that the new model stratified the risk of prostate cancer death significantly better than the widely adopted three-tiered classification of low, intermediate and high risk [10–13]. In this paper, we report validation of this model, called the Cambridge Prognostic Groups (CPGs), in two separate, ethnically different cohorts: 72,337 Swedish men and 2550 men from a Southeast Asian population. Using the Swedish study group, we also assessed the utility of the CPG model in pre-treatment prognosis in men who had surgery, radiotherapy or conservative management.

The CPG model, now tested in three different international cohorts in two studies including 86,732 primary prostate cancers, delivers distinct cancer mortality sub-groups with a high prognostic accuracy. The prognostic power of the model was very consistent between our development cohort and this validation study [13]. To our knowledge, the CPG model is the first risk stratification tool to have been derived from and validated in cohorts of newly diagnosed men using cancer death as the primary outcome. Our tested cohorts also included significant proportions of locally advanced cases (12 and 16%) and men managed conservatively (19 and 21%), which reflect most real-world practices where PSA screening is uncommon and unlikely to be implemented [5, 20–22].

The intermediate-risk group is the largest category of patients in contemporary cohorts [5]. The CPG model divides this group into two categories: CPG2, which is associated with a relatively good prognosis, and CPG3 (a combination of intermediate-risk factors or Gleason Grade Group 3 on its own), with a substantially higher mortality risk despite radical therapy. This data supports the recent work of Raldow et al., where men with multiple intermediate-risk features had higher rates of prostate cancer death following brachytherapy [23]. Our results further suggest that many men with CPG2 disease may potentially be candidates for conservative management, at least initially, and thus avoid the morbidity of unnecessary treatment. In contrast, men with CPG3 should not be managed conservatively, as they have a much higher baseline risk of PCM. We do interpret this with caution, as our data may be potentially biased by treatment selection. Nevertheless, our results are supported by the work of Musunura et al., who observed a similar worse survival outcome from active surveillance in men with a combination of Gleason 7 and a high PSA [24]. Our distinction between CPG2 and CPG3 has now also been independently identified by the new 2017 American Urological Association (AUA)/American Society for Radiation Oncology (ASTRO)/Society of Urologic Oncology (SUO) localised prostate cancer guidelines. They have defined a favourable and unfavourable category amongst intermediate-risk cancers, the criteria of which perfectly match the ones used here in the CPG2 and CPG3 categories [25]. Although the AUA/ASTRO/SUO definitions were not derived from primary research, they endorse our evidence-based distinction from a large cohort study that these sub-groups are linked to very different mortality outcomes. Consistently across all treatments, the split of the traditional high-risk category into CPG4 and CPG5 (multiple high-risk features) defined groups with very different risks of cancer death. CPG5 men had more than double the risk of PCM, even when compared to CPG4. These results support the findings of previous studies reporting that multiple high-risk factors confer a much worse treatment-specific outcome [26, 27]. Our study is, however, the first to show this effect in a very large cohort and simultaneously across different treatment types. Men with CPG3 and CPG4 disease represented statistically different prognostic sub-groups in our overall cohort analysis with distinctly different outcomes in intergroup comparisons. This mirrors the findings of our initial development study [13]. However, we do show for the first time that they may have very similar outcomes when treated by radical therapy. The reason for this may be that these treatment sub-cohorts were too small to pick up a difference, but our findings do support the notion that CPG3 likely represents a distinct aggressive sub-type of intermediate-risk disease more akin to the traditional high-risk disease designation.

A consistent criticism of risk and prognostic groupings is that they do not address intra-group heterogeneity [28]. As an example, Joniau et al. showed that amongst very high-risk men (in our study, CPG5) having surgery, the sub-group T3 and PSA > 20 had better outcomes compared to men with very high Gleason score 9–10 disease [27]. Although this criticism could also be applied to the CPG, we believe that our stratification system is an important first step in providing a more accurate but still simple framework for more individualised decision-making in non-metastatic prostate cancer. Hence, when we looked at our very high-risk group, the different categories did all have significantly worse mortality outcomes compared to CPG4 (the next prognostic level), p < 0.001 in all comparisons. In terms of practical usage, we believe that the CPG groups add significant clinical benefit. For example, men in CPG1 should be preferentially steered towards active surveillance. Many men in CPG2 are also likely to do well from this option but may need a more intensive surveillance schedule. In contrast, men in CPG3 and CPG4 clearly need curative therapy, and for these men the added use of individualised estimates of treatment outcomes could be very helpful. Bespoke biomarkers could also be used which are more appropriate for the disease context. A recent example is the work of Ahmad et al., who showed that adding a DNA methylation index to the Cancer of the Prostate Risk Assessment (CAPRA) score improved prediction of PCM in men with intermediate-risk disease (area under the curve (AUC) from 0.62 to 0.74) [29]. Fraser et al. also studied men with intermediate-risk disease having radical therapy and demonstrated the utility of a panel of 40 recurrent genomic alterations in identifying those at highest risk of treatment failure [30]. Hence in the future, improved outcome discrimination within the CPG sub-groups might be gained by including such factors to add granularity. Men in CPG5 in particular clearly need a more aggressive and new approach to treatment and may be the ideal cohort for molecular sub-typing and targeted neo-adjuvant drugs combined with radical therapy when planning new clinical trials [31]. Conversely, it is likely to be a waste of resources to do such profiling in men with already good outcomes (e.g. those in CPG1). The CPG model may also be used to construct tailored follow-up protocols. For instance, men with CPG5 disease are likely to benefit most from early adjuvant treatment after radical therapy compared to men with CPG4 because of a much higher risk of a poor outcome. Conversely, in a surveillance programme, men in CPG1 are likely to only need a low-intensity follow-up schedule. A trigger for conversion to treatment might then be an increment to a higher CPG category during follow-up evaluation.

Our study does have limitations. It has been built and validated on men who have been diagnosed via trans-rectal ultrasound-guided biopsy, which is known to underestimate true histological grade and overall tumour burden [32]. However, the contribution that more intensive biopsy schema might make is currently uncertain. The ProtecT Study, for instance, showed extremely low mortality rates at 10 years in the surveillance cohort, despite the fact that men only had this kind of biopsy and at least a third likely harboured missed higher risk disease [33]. We did not have data on biopsy core involvement in our cohorts, and it was not a requirement for our model; thus, we cannot say if such granularity would improve its prognostic power. We note that biopsy core involvement is not included in contemporary guidelines outside the USA, and there is no international consensus on its use [10–12]. We also did not sub-classify within T stages, but we have previously noted the inaccuracies in its standard clinical use [34]. Our cohort predates the use of magnetic resonance imaging (MRI) for guided biopsies, which is already changing clinical practice [35, 36]. The CPG model, however, will retain utility regardless of the biopsy approach, as it is based on standard clinico-pathological variables. Indeed, we have already demonstrated the use of the model with MRI-based staging in predicting bone metastasis at diagnosis [37]. About 11% of the PCBaSe cohort had to be excluded, as we did not have all the clinico-pathological details. Details of how missing data is handled in PCBaSe have been previously reported [38]. Finally, although we have included competing mortality risks, our model does not include co-morbidity as a variable. Indeed, none of the current UK and European prostate cancer guidelines do so [10–12]. The US National Comprehensive Cancer Network (NCCN) guidelines also only go as far as to distinguish a life expectancy of less or more than 5 years [39]. The CPG model can of course be used alongside other tools to predict other-cause mortality [40].

In summary, here we have confirmed the superiority of a five-tiered prognostic system over the prevailing three-tiered model in better stratifying prognosis in men with non-metastatic prostate cancer. Our model is unique in that it has been built from primary diagnostic source data linked to PCM outcomes, it has been tested across different treatment types and now has been validated in very large cohorts of men. It is notable that in the USA the NCCN and new AUA/ASTRO/SUO localised prostate cancer guidelines are both also endorsing five-strata systems for non-metastatic disease although with different sub-categories [25, 38]. While the AUA/ASTRO/SUO endorse splitting the old intermediate-risk category, they do not recommend a very high-risk sub-category. The NCCN criteria, in contrast, endorse a very high-risk category but do not endorse splitting the intermediate-risk group. The CPG model underscores the clinical relevance of sub-stratification of the old intermediate-risk group in the new AUA/ASTRO/SUO guidelines as well as the designation of a very high-risk category but combines these two new sub-groups in the only single, easy-to-reference, evidence-based model. While the prognostic performance of the CPG model has remained consistent in three cohorts from different countries, future independent validation would strengthen its clinical utility. Our model, however, can be used by any clinician anywhere in the world without requiring any additional data or costs. Future work will determine how additional variables, including biopsy data and molecular profiles, can be added to further individualise prognostic prediction both across and within treatment sub-types.