Optimizing the clinical utility of PCA3 to diagnose prostate cancer in initial prostate biopsy

Several nomograms optimize the indication of IBx, with predictive accuracy estimates between 67 and 77 % [3, 24‐26]. In particular, the PCA3-based nomogram validated here showed that the accuracy of the clinical model was increased by 4.5–7.1 % related to PCA3 inclusion [15]. In this model, continuously coded PCA3 represented the most informative parameter in the prediction of any PCa (AUC = 0.739) and HGPCa (AUC = 0.729). In our univariate analyses, we also showed that continuous coded PCA3 was the best predictive variable for the detection of any PCa (AUC = 0.701) and prostate volume was slightly better to detect HGPCa (AUC = 0.723) than continued coded PCA3 (AUC = 0.685). Checking different cut-offs, 24 and not 21, showed the highest AUC for both any PCa and HGPCa (Table 2). These findings agree with previous studies [27]. We also tested categorized PCA3, and we agree with the Hansen et al. series that continuous coded PCA3 was the most informative variable to predict PCa and HGPCa, therefore it should be considered this way when building a nomogram. The non-linear effects of PCA3 can be appropriately modeled using its continuous form, thus adding more predictive power [28].

The calibration plot and the values of intercept and slope showed a poor agreement between actual and predicted probabilities obtained from the application of the nomogram to our 401 men as a validation cohort (Fig. 1). This fact is explained by the clear differences in detection rates between both series. Our PCa detection rate was 28 % (11 % HGPCa), closer to the expected prevalence of PCa in contemporary series of opportunistic screening, and the same features were 46 % (20 % HGPCa) in the referenced series. The initial cohorts which generated Hansen’s nomogram consisted in 570 patients enrolled from 4 North American sites with a detection rate of PCa of 36 % (15 % HGPCa) [17] and 516 patients from a European multicenter study with a detection rate of PCa of 40 % (19 % HGPCa) [4], closer to our data. From the entire cohort of 1086 patients, 692 patients were finally considered to generate Hansen’s nomogram and the proportion of PCa and HGPCa increased substantially (46 % PCa and 20 % HGPCa) from original cohorts. We think this fact could explain their high prevalence of PCa and HGPCa, which overcomes expected rates of PCa and HGPCa for patients with PSA between 2,5 and 10 ng/mL. In this sense we think our detection rates are more similar to the reported in the literature, with most men coming from opportunistic screening scenarios. Only series with so extremely high rate of PCa/HGPCa could expect a nice calibration plot. In any case, a good calibration would had shown a good performance of the Hansen et al. model in our series, but a bad calibration doesn’t mean that the validated nomogram is a bad predictive model in our series, but just a different threshold point must be investigated.

We obtained slightly inferior AUC (0.736, CI95 %:0.680–0.793) for any PCa detection that the referenced nomogram (AUC 0.807, CI95 %:0.768–0.828), showing statistically significant differences between AUCs (p-value = 0.02), but similar to the other published external validation (0.764, CI95 %:0.726–0.802) [27], p-value = 0.43, so we think the model offers good discrimination ability. When we built a logistic regression model using the same predictive variables and using PCA3 as a continuous value we obtain an AUC of 0.769 (data not shown), very similar to the application of Hansen nomogram in our series (Table 3). Recently, a similar evaluation of the use of PCA3 as a continuous predictor in a multivariate logistic regression model developed over 3073 patients from screening population in USA, showed an AUC of 0.75 and 0.81 respectively for the prediction of PCa and HGPCa respectively at the IBx [29].

In the decision curve analysis we obtained at a 30 % threshold probability a net benefit 8.41 %, superior to the baseline model, but far away of the >18 % recognized for the model in the original series. For a threshold probability of 40 % the net benefit obtained is 6.31 %. Other authors have also shown that decision curve analysis confirmed a higher benefit when adding the PCA3 score (either continuous or binary with a cutoff of 35) to the baseline model [27] in IBx. In the logistic regression model built with our database using the continuous PCA3 score, the net benefit is 6.90 % for the 40 % threshold point, very similar to the Hansen nomogram application.

Multivariate models translate multiple effects in one number, which is the interpretation of the risk of harboring an event from 0 to 100 %. But we as clinicians take decisions based on a reference value (PSA > 4 ng/ml, free-PSA ratio < 15 %, etc.), counseling patients taking based on them. With the aim of helping the clinician to indicate or not an IBx, we investigated the probabilities of the model to detect PCa through PDF [16]. These density functions help us to choose thresholds to differentiate groups of high and low risk probabilities of harboring PCa (Fig. 3a) or HGPCa (Fig. 3b). The density curves of probabilities are built from the probabilities of having PCa assigned by the nomogram to each man. In their X-axis, we reflect the range of possible probabilities (0–100) of harboring PCa according to the nomogram. In the Y-axis we range the density of patients from their assigned probability of harboring PCa or HGPCa. We can see the higher peak of density in PCa patients near the X = 70 % probability of harboring PCa in Fig. 3a and in HGPCa patients near 80 % in Fig. 3b. On the other hand, this peak appears around X = 15–20 % for non PCa/non HGPCa patients.

We can easily see how our PDF curves show a “valley” between the “peak” of patients with PCa/HGPCa in the range of high probabilities of harboring PCa (in the right side of the graphic), and the “peak” of patients without PCa/HGPCa in the range of low probabilities (in the left side). This valley drives us to choose the threshold of probability to classify patients in high or low risk of harboring PCa. A threshold point of 40 % instead of 30 %, as proposed by Hansen et al., could be the best option to translate the implementation of this nomogram in our daily practice.

Applying the nomogram with a threshold of 40 % to our 401 men, we would had saved 197 IBx (49.1 %), at a cost of missing 27 any PCa (24.3 %) and 7 HGPCa (15.9 %) (Table 4). For the threshold value of 30 % provided by Hansen et al., we would had saved 151 IBx (37.3 %) at a cost of missing 21 any PCa (19.9 %), and 6 HGPCa (13.4 %). Therefore, selecting 40 %, nor 30 %, we would had saved 11.8 % IBx more just missing one HGPCa more. This features would always improve the results of taking single PCA3 cut offs values as a single tool to decide IBx (Table 5), where we can check that if we had chosen PCA3 > 21 we would had missed 15.9 % HGPCa, similar to the 13 % observed by other authors at IBx with a cut-off of 20 [30], but doing 12.2 % more IBx that if we had applied the nomogram. We notice that our small number of HGPCa (44 cases) could affect our data on this population, as using the threshold point of 40 % the percentage of missed HGPCa cases is 15.9 %, but the 95 % CI is 7.1–30.7 %.

It would had been desirable to compare the initial clinical nomogram built without PCA3 evaluated by Hansen et al. to ours, in order to know the clinical benefit of determining PCA3, but that nomogram was not published. We show in Table 3 that there are statistically significant differences (p < 0.01) between models build with or without PCA3 as predictor variable.

From a practical point of view, and in order to save costs, we ask for PCA3 just in doubtful cases, in the way Abern and Freedland propose [31]. If we had applied the nomogram to our series, we would had obtained a score of 121 total points (equivalents to a probability of ≥ 40 %) in 204 men. Twenty-six of them would had had 121 points without the need to test PCA3, so we would had indicated the IBx saving costs. In the lower scenario, 178 would not had reached to 121 points adding the additional 26 points dependent on a PCA3 > 21, but we think that not using the aid of PCA3 at this scenario, knowing the strength of PCA3 as a continuous variable and that prostate volume could be undermeasured by hypogastric sonography, that the cost of PCA3 would be worth while for a better counseling of IBx to a men in this grey area.

Finally, this external validation in a single center over a series of 401 patients is closer to a opportunistic screening scenario with a prevalence of PCa of 28 %, more common than the 46 % given by the referenced nomogram. This fact makes it particularly applicable in daily practice compared to the referenced nomogram (international, multicentre, multiethnic, different PSA assays used).