Five UK centres took part with each recruiting consecutive men referred from primary care for elevated prostate-specific antigen (PSA) (January 2018 to June 2019.) mpMRI method and reading and biopsies were carried out according to local protocols in order to ground the study in real-world practices. The only stipulation was image-guided targets + systematic biopsies (cognitive or image fusion). Men were excluded if they had (i) a previous biopsy, (ii) pelvic metalwork interfering with mpMRI quality or no mpMRI and (iii) if no biopsy was done after mpMRI. PSA and phi assay was taken before biopsy and PSA density (PSAd) calculated using mpMRI-defined prostate volumes. Only men with intact information on key data points, PSA (ng/mL), phi, mpMRI, prostate volume, biopsy performed, histological Grade Group and clinical tumour stage (T-stage) were included in the final analysis. Men with missing data from the original 554 were not included in the analysis leaving a final study cohort of 545 (Additional File Fig. S0). Additional File Table S1 details the imaging and biopsy strategy in each unit. Men with mpMRI-negative lesions had systematic sectoral biopsies only. The study was conducted under ethics REC 03/018.

The primary end-points were two definitions of prognostically important cancers (i.e. cancers that may shorten life-expectancy if not found). For this, we used International Society of Urological Pathology histological Grade Group 2 or more on biopsy (≥ GG2) and prognostic group 3 or more using the composite Cambridge Prognostic Group (CPG) prognostic score (≥ CPG3) we have previously reported and validated [22]. CPG is a five-tiered multi-factor (PSA, Grade and Stage) prognostic model for non-metastatic prostate cancer shown to have superior discrimination in predicting prostate cancer deaths compared to any other tiered stratification systems [23]. Disease ≥ CPG3 is similar to unfavourable intermediate risk and high-risk disease in the AUA, NICE and EAU systems [24]. Areas under the receiver-operating curve (AUC) and diagnostic test statistics for different strategies were compared to PSA alone: phi and PSAd. We adopted a decision modelling approach to bring together information about sensitivity, specificity and costs. A range of potential clinical pathways were modelled and compared Strategy 1: mpMRI and biopsy all, Strategy 2: mpMRI all and biopsy if positive, Strategy 3: mpMRI all and biopsy if PSAd ≥ 0.15, Strategy 4: mpMRI all and biopsy if PSAd ≥ 0.1, Strategy 5: phi all and mpMRI and biopsy if phi ≥ 25, Strategy 6: phi all and mpMRI and biopsy if phi ≥ 30. The proportions of men with positive test results in the study cohort were used to calculate cancers detected or missed under each pathway. Additional Figure S2 shows the pathway and cohort percentages corresponding to Strategy 6. Costs for assays, scans and biopsies are given in Supplement Table S2; one-way sensitivity analyses were used to explore the impact of assumptions regarding the cost and risk of sepsis following biopsy. Key outcomes relating to correctly identified cancers (true positives), incorrectly identified cancers (false positives), missed cancers (false negatives), numbers of MRI scans, numbers of biopsies and total costs were then calculated for a hypothetical cohort of 1000 referred individuals.

Decision curve analysis (DCA) was done to estimate the clinical net benefit outcome, which accounts for the perceived value weighting between the harms of biopsy in those without cancer and the harms of not identifying those with cancer [25]. The value weighting represents the decision-makers’ beliefs about the benefit to harm ratio between cancer detection and unnecessary biopsy. It can be thought of as the risk of cancer at which there is equipoise about whether to proceed to biopsy (risk threshold, RT) and takes values between 0 and 1. A low risk threshold implies that the perceived harm of biopsy is low compared to the benefits of identifying cancer; conversely, a high-risk threshold occurs when biopsy harm is high compared to the benefits of identifying cancers. The clinical net benefit is then calculated (at a given value of RT) as the number of cancers detected minus the number of unnecessary biopsies weighted by RT. Since the perceived point of biopsy equipoise is uncertain, results are presented over a range of RT values. The preferred detection strategy is then indicated by the one with the highest net benefit, at a given RT. The cost per net cancer detected under each strategy is also presented.

The final study population included 545 men with a median age of 66 years, PSA of 8 ng/ml and phi of 44 (Table 1). Overall mpMRI results were reported as positive (M3-M5) in 420/545 (77%) with M4-M5 lesions found in 316/545 (58%) (Table 1). The median number of cores taken was 16 for systematic biopsies and 2 for targets (if positive on an mpMRI). Overall, prostate cancer was detected in 349/545 men (64%) and ≥ GG2 cancers in 256/545 (47%). Using the composite CPG score, disease ≥ CPG3 was found in 174/545 men (32%). Additional File Table S1 details the MRI positivity and method of biopsy acquisition for each centre. There were no significant differences between centres in any of these parameters despite no pre-specified standardisation for diagnostic method and reporting [26, 27].

Descriptive characteristics of the primary study cohort. MRI positive data is shown as the PI-RADS score of ≥ 3 or ≥ 4. Detection rates for cancer are shown for any cancer, and using definitions of ≥ Grade Group 2 (GG2 or ≥ Cambridge Prognostic Group 3 [CPG3]. Excludes 9 men with missing data not included in the analysis

For detection of significant cancers defined as ≥ GG2, the AUCs were 0.70, 0.79 and 0.82 for PSA, PSAd and phi respectively (Fig. 1a and Table 2). Both PSAd and phi performed significantly better than PSA alone in predicting ≥ GG2 disease (p < 0.001) (Table 2). The predictive value of phi for ≥ GG2 was also similar across all 5 centres (p = 0.67). The combination of PSA, PSAd with imaging increased the AUC to 0.76 and 0.81 respectively (Table 2). Using ≥ CPG3 as an endpoint, the AUCs were 0.81, 0.84 and 0.87 for PSA, PSAd and phi respectively (Additional File Figure S1A and Table 2). However, only phi was significantly better than PSA in predicting ≥ CPG3 disease (p < 0.001) (Table 2). Amongst mpMRI-negative men (n = 125), the AUCs for detection of ≥ GG2 cancers were 0.64, 0.76 and 0.78 respectively with both PSAd and phi performing better than PSA. Neither PSAd nor phi however performed better than PSA in predicting the presence of ≥ CPG3 disease (Fig. 1b, Table 2 and Additional File Figure S1B). Fig. 1

ROC curve illustrating performance of phi, PSA, PSAD and mpMRI in predicting cancer diagnosis of ≥ Grade Group 2 (GG2) in the a whole cohort and b mpMRI-negative men (PI-RADS ≤ 3)

We next tested different threshold of the phi to test its ability to triage men for initial mpMRI and biopsy. Using detection of ≥ GG2 as an endpoint, phi cut-offs ≥ 20 and ≥ 30 to refer for mpMRI and biopsy had an NPV of 0.85 and 0.90 respectively and missed 1.1 and 7.7% of tumours (Table 3). With ≥ CPG3 as a detection target, phi performance showed even better results with NPVs of 0.94 and 1.0 respectively and with 0% and 4.5% of cancers missed. Of note, using a phi ≥ 35 threshold (which we reported for re-biopsy men) produced rather poor sensitivity [21] (Table 3). This suggests that useful phi thresholds may be different depending on the detection context. As a comparator, we also analysed the performance of PSAd (noting that this is not usually available before mpMRI). PSAd thresholds of ≥ 0.1, ≥ 0.15 and ≥ 0.2 missed between 7 and 31% of ≥ GG2 cancers and 3–20% of ≥ CPG3 tumours (Table 3).

Diagnostic test statistics of the study cohort showing the accuracy and missed cancer rates for each phi threshold as a triage test pre-MRI and biopsy. PSAd thresholds are shown as a comparator though this metric is not usually available before an mpMRI. Detection rates for cancer are shown using definitions of ≥ Grade Group 2 (GG2) or ≥ Cambridge Prognostic Group 3 [CPG3]. (*percentage out of 258 cancers detected, ** percentage out of 176 cancers detected)

We next compared different pathways to model referrals with and without use of phi. Table 4 summarises the key modelling results for each strategy, based on a hypothetical cohort of 1000 men referred for suspected prostate cancer. The base case model was an mpMRI and biopsy for all which detected every cancer but required more than half of men (53%) to undergo unnecessary biopsies (i.e. benign histology or ≤ GG2) (Table 4). Each alternative option resulted in an overall reduction in biopsy procedures, with the most impactful being a strategy to only biopsy men with positive mpMRI (M3–5) or to only biopsy men with a phi of ≥ 30 (23% and 25% reduction respectively). phi ≥ 30 also achieved this with a concomitant 25% reduction in mpMRI use by virtue of being an upfront triage test. We next assessed the impact on rates of detection of significant cancers. Using detection of ≥ GG2 as an endpoint, a strategy of mpMRI and biopsy only positive cases reduced unnecessary procedures by 35% but did miss 9% of tumours. In contrast, the phi ≥ 30 option to triage in referrals reduced unnecessary biopsies by 40% and missed 8% of ≥ GG2 cancers. The phi ≥ 25 option missed even fewer ≥ GG2 cancers (4%), but only reduced unnecessary biopsies by 25%. Using detection of ≥ CPG3 as an endpoint, both strategies of mpMRI and biopsy-positive lesions and using phi ≥ 30 led to similarly low missed cancer detection rates (5%); however, the phi ≥ 30 route required fewer unnecessary biopsies (31% vs 35%). Other pathways using PSAd (≥ 0.10 and 0.15) required many more overall biopsies (reductions of only 6–12%) and by definition all needed mpMRI (Table 4). These strategies did though miss many fewer ≥ GG2 tumours (1–2%) and no ≥ CPG3 disease (Table 4).

Modelling results for a hypothetical cohort of 1000 patients referred for elevated PSA using different strategies for cancer detection. In brackets are projected reductions compared to the base model (* compared to MRI + biopsy all)

Table 4 includes the mean cost per investigated patient for each of the modelled diagnostic strategies. Compared to a cost of £965/person (mpMRI + biopsy) for all referred men, using a phi ≥ 30 threshold to restrict investigations reduced costs by 20% (£774/person). In contrast, a strategy of mpMRI + biopsy for only scan-positive cases was marginally costlier (£796/person). Strategies that involved using PSAd were also more expensive mainly due to the increased use of both mpMRI and biopsy (mean cost £879 and £921 for PSAd ≥ 0.15 and 0.10 respectively). Sensitivity analyses using higher rate or cost for sepsis following biopsy showed little impact on these comparisons (results not shown). Figure 2 shows the results of the decision curve analysis (incorporating the perceived harms of biopsying those without cancer) in terms of net clinical benefit resulting from different strategies based on ≥ GG2 tumour detection. At very low values of the risk threshold, biopsy all is the optimal strategy because these values represent a belief that there is little to no harm associated with an unnecessary biopsy. For risk thresholds 0.2–0.5 (the estimated range of uncertainty in clinical practice), the net benefit is highest under the phi ≥ 30 pathway. Since this is also the cheapest option (Table 4), this appears to be the optimal testing strategy representing a cost/cancer detected of between £2120 and £5860 (Additional File Table S3), depending on the risk threshold. However, for risk thresholds < 0.2 (i.e. where there is uncertainty to biopsy at cancer risk < 20%), the clinical net benefit is maximised by using more costly strategies. The optimal decision at these lower-risk thresholds therefore depends on the willingness to pay/cancer detected (Additional File Table S3). Hence for risk thresholds < 0.2, the cost for the phi ≥ 30 strategy is around £2000/cancer detected but more cancers (net) could be found using other strategies, at the cost of £4000 to £8000/additional cancer depending on the pathway. Fig. 2

Decision curve analysis comparing the number of net benefits for detection of significant cancers for a range of risk threshold values and using different approaches (MRI-PSAd using a PSAd threshold of ≥ 0.15). MRI - magnetic resonance imaging, PSAd - PSA density, phi - Prostate Health Index