Prostate cancer represents a growing burden on health care globally, with increasing numbers and proportions of men presenting with non-metastatic prostate cancer (PCa) [1]. Alongside this, there has been increased confidence in the use of conservative management (active surveillance and watchful waiting) [2]. Understanding disease prognosis to guide treatment decision-making is therefore of great importance. However, until recently, no high-quality individualised model for survival existed.

Using data from over 10,000 UK men, we have previously published an individualised prognostic model for cancer-specific and overall survival called ‘PREDICT Prostate’ [3]. PREDICT Prostate (available online [4]) provides cancer-specific and overall percentage survival estimates for up to 15 years and has been endorsed by the National Institute for Health and Care Excellence (NICE) [5]. To maximise usability, it uses routinely available clinico-pathological data (age, PSA, grade, stage, biopsy involvement, treatment type and comorbidity). It represents real-world data from a non-screened, primary diagnostic cohort, including a significant number of men treated conservatively. Crucially, the model also allows adjustment for competing mortalities by incorporating both cancer-specific and non-cancer survival outcomes to contextualise the diagnosis as part of a decision aid. Internal validation and accuracy within a small external population were promising during model development [3]. However, external validation in independent cohorts, ideally in a different location, is vital to demonstrate generalisability and accuracy of a multivariable prognostic model [6].

The Prostate Cancer data Base Sweden (PCBaSe) is one of the largest and most comprehensive prostate cancer cohorts world-wide and is well suited for external validation of PREDICT Prostate [7]. The aim of this study was to validate PREDICT Prostate and compare performance to existing models.

In this large external validation cohort, we demonstrated that PREDICT Prostate is a robust and generalisable long-term prognostic model. In the analysis of an independent cohort, ten times larger than the original cohort, discriminatory accuracy and calibration was good. This also remained true within treatment groups, particularly in men managed conservatively or by radical therapy.

Conveying information to an individual about their disease prognosis within their own context of competing mortality has historically been an imprecise exercise with little objective data available. The most current prognostication is based on stratification groups of the cancer itself and discussions with clinicians who may be conflicted towards a particular treatment [19–21]. PREDICT Prostate was conceived to address this gap in clinical need and standardise the decision-making process [3] and has shown promise to positively influence clinical decision-making [22]. It is built around long-term actual survival data and has been designed to address all AJCC criteria [6].

In the model development study, c-indices were 0.84 for PCSM and 0.77 for ACM within the UK validation cohort [3]. In the original study, external validity was also assessed within a Singaporean cohort. However, this cohort was small (n = 2546) and follow-up was quite short (5.1 years). Here we show in a cohort of > 69,000 men with longer median follow-up that our c-indices were actually improved to 0.85 for PCSM and 0.79 for ACM with excellent calibration. We did note a marginal overestimation of PCSM, which was contrary to the slight underestimation we had observed in the Singapore external validation in the original paper [3]. Given that the model is very well calibrated for ACM, this apparent overestimation of PCSM (and corresponding underestimation of NPCM) is likely to be a result of differences in cause of death classification, reporting or recording practices. ACM is the key outcome of interest, and a more unequivocal endpoint, against which this model performs very well.

When compared to existing models, PREDICT Prostate consistently outperformed the three-stratum risk classification system used in the EAU, D’Amico and NICE stratification criteria [16, 20, 23]. We recognise that comparisons against these risk stratification criteria are limited and that they are not designed to be prognostic nomograms; however, they are widely used in clinical practice to inform treatment decisions. Benefits of PREDICT were also seen against the NCCN and CAPRA scores, which add more granularity but ultimately retain a grouping system rather than individual estimates [16, 17]. For the outcome of PCSM, the CAPRA score did perform similarly well for some treatment groups, particularly in men treated with prostatectomy. This is unsurprising, as the model was originally built around prostatectomy patients [24]. It should be noted that PREDICT Prostate is not a treatment-specific tool; therefore, by assessing discrimination within treatment sub-groups, its discriminatory performance will inevitably be reduced. Nonetheless, PREDICT Prostate performed significantly better in predicting ACM and PCSM in most treatment groups. We also confirmed that adding in biopsy data to the model improved the performance though this effect was marginal in addition to the other variables already included. Using PPC as a continuous variable maximises the use of prognostic information, and this parameterisation did lead to marginally superior discrimination for ACM.

The primary utility of PREDICT Prostate will be in men for whom conservative management and radical treatment might both be appropriate options, for whom the decision is most difficult. Abundant literature demonstrates that decision aids contribute to more knowledgeable and informed patients and that they can improve clinician-patient communication [25, 26]. Therefore, the model may have wide potential applications in informing patient, clinician and multi-disciplinary team decision-making to reduce both over- and under-treatment. Formal clinical impact assessments are also crucial to show face and functional validity, and these are underway with PREDICT Prostate [27]. Future research endeavours could assess what impact the use of the model might have on actual treatment practices and compare this model with prognostic biomarkers, or radiological prognosticators. Over time, additional parameters can be incorporated into this base model, or the model itself be updated, should new variables be shown to have independent prognostic effects [28].

More recent efforts in prognostic tools have sought to utilise novel genomic or biological markers to generate prognostic estimates. However, most established genomic tools such as Prolaris CCP and Oncotype DX GPS have predominantly been tested against shorter-term outcomes or in treatment-specific cohorts [29, 30]. Where they have been assessed against PCSM, concordance has been very similar to our model—for example, the Decipher genomic classifier alongside CAPRA showed an AUC of 0.78 (95% CI 0.68–0.87) for 10-year PCSM following prostatectomy [31]. Direct comparison with PREDICT Prostate is not possible without a head-to-head or combined study, but the value of such expensive tests do need to be re-assessed in the context of optimised clinical multivariable models [32]. In this context, we would welcome collaborations or independent studies on the value of adding genomic classifiers to future iterations of PREDICT Prostate.

This study has numerous strengths, given the large sample size, long follow-up and high completeness of data in PCBaSe [33]. However, we recognise limitations inherent to using registry data. Seventeen per cent of men were excluded due to missing data, and we cannot exclude this, introducing some bias. A large proportion of men within this validation dataset had low-grade disease, such that PCa mortality rates were relatively low which may affect discriminatory performance. Men diagnosed within the inclusion period may also not be representative of contemporary practice with changes in PCa diagnosis and treatment. For instance, we recognise that primary hormone therapy is now rarely used in the context of non-metastatic PCa; hence, we included sub-group analyses within other treatment groups. We also appreciate that multi-modal therapies are increasingly used in higher risk cases, which we were not able to assess in this study due to the inclusion dates and data availability limitations of our datasets. Another particular concern is the lack of information from magnetic resonance imaging (MRI). However, the current focus for MRI is on tumour detection rather than prognostication and it is unknown if MRI lesion characteristics (Likert or PIRAD scoring) have any bearing on survival. Our model also cannot account for subsequent transitions to different treatments. However, in our UK dataset, conversions to active treatment were less than 6% across total follow-up [3]. We also recognise the lack of T-stage sub-classification, which is a key parameter in 2 of the existing models we made comparisons to. However, it is accepted that T stage is often inaccurately assigned in localised disease [34]. We also recognise that other endpoints of interest exist, particularly the development of metastases and commencement of hormone therapy. The model is untested against these endpoints, but calibrated against the more robust endpoint of death.

A key issue going forward is the validation of this model in non-Caucasian and screened populations. Although the original paper re-tested the model in Singaporean men, PREDICT Prostate remains untested in men of African descent or other ethnicities. Independent validations within screened populations, and within other prospectively collected or randomised datasets, would also be helpful and should be encouraged. Finally, we recognise that other nomograms are available, against which direct comparisons would be very insightful. These were not possible within the design of this study, or the limitations of this data, particularly with regard to comorbidity.

This large external validation demonstrates the robustness of PREDICT Prostate. PREDICT Prostate, available as a free-to-use web tool [4], has the potential to significantly improve shared decision-making for PCa management, particularly the choice between conservative management and radical treatment. Further, independent external validations are encouraged, especially in populations of different ethnicities.