The institutional review board of two hospitals approved this retrospective study from a prospective cohort of mpMRI fusion-targeted biopsies (14) and waived the requirement for informed consent. All patients in this study underwent prostate mpMRI and transperineal systematic plus targeted prostate biopsy under local anesthesia sequentially from October 2016 to February 2020 in Drum Tower Hospital and Molinette Hospital, as previously described (13, 14). Those who were diagnosed with PCa and subsequently treated with robot-assisted laparoscopic RP (RARP) were included. More details about the criteria are shown in Figure 1 .

All patients underwent mpMRI performed with a 1.5- or 3-Tesla system (details shown in Supplementary Table 1 ) (15, 16). Two radiologists with over 10-year experience in urology image analyses supervised the results using the Prostate Imaging-Reporting and Data System (PI-RADS) v2 standards (17). All radiologists had at least 1,000 prostate MRI images of reading experience. The protocol consisted of T2-weighted (T2W) imaging in three planes, diffusion-weighted imaging (DWI) with the calculation of apparent diffusion coefficient (ADC) maps, high b-value images (b > 1,500 s/mm²), and dynamic contrast-enhanced imaging (18). In case there was disagreement over the outcome of the image, the radiologists would discuss it until they reached a consensus.

Biopsy targets were defined as regions of interest (ROIs), which were lesions with PI-RADS score ≥ 3. Before the biopsy, operators annotated targets using an ultrasound system (Esaote Real Time Virtual Sonography, Hitachi Medical Corporation, Tokyo, Japan) with reference to the reports from the two radiologists. All patients underwent transperineal MRI–ultrasonography (MRI-US) fusion biopsy (18G needles with sampling length 17 mm) with a US diagnostic system (MyLab Twice, Esaote S.p.A., Genoa, Italy) consisting of 12-core SB and 2- to 4-core targeted biopsies for each ROI under local anesthesia, as previously described (13, 14).

Histopathology of prostate biopsies was performed by specialized urological pathologists independent of MRI results. For cores with cancer, grade group (GG) and Gleason score (GS) were determined by using the 2014 International Society of Urological Pathology (ISUP) guidelines (19). csPCa was defined as different criteria according to MRI-FIRST study (csPCa-1, ISUP GG 2 or higher tumors; csPCa-2, ISUP GG 3 or higher tumors) (20).

Retrospective collection of urology, radiology, and histopathology data for all patients included the following clinical characteristics: histopathology result of biopsy and RP, age, height, weight, body mass index (BMI), pre-biopsy serum prostate-specific antigen (PSA) value (ng/ml), prostate volume, the maximum diameter of lesion, gap days between biopsy and RP, number of cores, and PI-RADS. The length of the prostate on anterior–posterior (AP), head–foot (HF), and right–left (RL) directions were also calculated on T2W mpMRI images.

“Upgrade” was defined as ISUP in RP pathology higher than ISUP in biopsy pathology, while “downgrade” was defined as ISUP in RP pathology lower than ISUP in biopsy pathology. These data were visualized under different biopsy situations including SB only, TB only, and SB combined with TB. At the same time, the upgrading and downgrading of different ISUP scores were also visualized. All the visualizations were conducted in Python v3.6.5 (Python Software Foundation) along with the machine learning classifiers with 10-fold cross-validation.

The machine learning algorithms were written using the Python SciKit-learn library except for eXtreme Gradient Boosting (XGBoost), which was written in an individual library called “XgBoost” (21). The logistic regression implemented in SciKit-Learn library was regularized logistic regression. Four classifiers were used including two base classifiers and two ensemble learning-based classifiers. In all classifiers, the training set was 395 of 565 (70%) patients, and the test set was the remaining 170 patients in the two institutions. Among the whole dataset, discrete variables were input to the one-hot encoder, denoting values taken on by categorical (discrete) features. The output would be a sparse matrix where each column corresponded to one possible value of one feature.

Firstly, logistic regression, one of the most classic classifiers to estimate the probability that a patient would have a particular outcome on the basis of related information or clinical characteristics, was used (22). Secondly, a support vector machine (SVM), a mathematical entity for maximizing a particular mathematical function with respect to a given collection of data, was used. The kernel function was used to “space up” the data to higher dimensions, which allowed the data to be linearly separated (23).

A scalable end-to-end tree-boosting system called XGBoost was then used, which could be built with a smaller sample size than most existing systems for ensemble learning-based classifiers. XGBoost is based on gradient tree boosting, an algorithm with which new models are created that predict the residuals of prior models and are then added together for the final prediction (24).

Finally, a random forest within a multiple decision tree model was used (25). Each tree was developed from a bootstrap sample of the training dataset, and each node was the part that was the best among a haphazardly chosen subset of features. The class predictions created by each tree within the forest were amassed, and the ultimate prediction was based on the lion’s share vote (26).

For all classifiers, we needed to choose optimal parameters based on the training set. A grid search was applied to iterate through each parameter combination. The best parameter set was confirmed by first plotting the receiver operating characteristic (ROC) curve and then selecting the one with the maximum area under the ROC curve (AUC). All classifiers were used to make predictions on the test set.

The performance of these classifiers was mainly evaluated by using overall accuracy, sensitivity, specificity, and AUC. Overall accuracy was the total correct ratio on all test sets. Sensitivity, similar to recall, indicated the proportion of upgraded PCa patients correctly identified by the classifier. Specificity indicated the proportion of patients who were correctly classified as “no-upgrade.” AUC indicated the probability that the classifier would have a higher prediction between “upgrade” and “no-upgrade” cases. All randomizations involved a random seed number of 0.

To evaluate the importance of these features and to figure out the most contributed features, the Gini index was leveraged. In the random-forest classifier, every time a node was split on a single feature, the Gini impurity criterion for the two descendent nodes was less than that of the parent node. Adding up the Gini decreases for each feature over all trees in the forest gives fast feature importance, defined as follows:

In the formula, K represents the total number of categories and p_k the probability that a case is divided into k categories in the case of a single feature (16). Our study was a binary classification, so K was equal to 2.

A nomogram was proposed to visually predict the risk of upgrading. A nomogram was composed of graphical lines of risk factors, points, total points, and upgrading probability. A calibration plot was used to validate how well the nomogram was calibrated. The x-axis of the calibration plot was defined as predicted probability, and the y-axis was defined as the actual probability (27).

Predicted probability was a prevalence rate calculated by a proposed nomogram. Actual probability could be calculated by dividing the number of patients with the same predicted probability by the total number of patients. An ideal line was drawn at a 45° angle in the calibration plot. The nomogram and calibration plot were plotted with R version 3.6.0 in package “rms” version 5.1-4.

From October 2016 to February 2020, a total of 694 patients underwent RARP after prostate biopsy in Drum Tower Hospital and Molinette Hospital. Of these patients, 179 were excluded because they did not meet the criteria. The remaining 515 were included ( Figure 1 ). Clinical characteristics of patients are summarized in Table 1 .

Regardless of the combination of SB and TB, of these 515 patients, only 245 (48.15%) cases were in accord with postoperative whole-mount histopathology, accompanied by 120 (23.30%) cases upgrading and 147 (27.18%) cases downgrading ( Figure 2 and Table 2 ).

The combined biopsy had a better concordance of postoperative histopathology than SB only (48.15% vs. 40.19%, p = 0.012). But there was no difference between the combined biopsy and TB only (48.15% vs. 42.33%, p = 0.069). The combined biopsy could significantly reduce the upgrading rate of postoperative pathology, in comparison to SB only (23.30% vs. 39.61%, p < 0.0001) or TB only (23.30% vs. 40.19%, p < 0.0001). Meanwhile, the combined biopsy would lead to a higher postoperative pathology downgrade rate (27.18% vs. 20.19% in SB, p = 0.002; and 27.18% vs. 17.48% in TB, p < 0.0001) ( Table 2 and Supplementary Figures 1 , 2 ). The pathology detail of primary and secondary Gleason patterns from SB and TB to RP is shown in Supplementary Table 2 .

Under different ISUP guidelines, 65 (53.28%), 29 (20.42%), 14 (11.29%), and 12 (10.81%) patients with ISUP1, ISUP2, ISUP3, and ISUP4, respectively, had upgrades as compared with postoperative whole-mount histopathology. At the same time, ISUP2, ISUP3, ISUP4, and ISUP5 had 8 (5.63%), 54 (43.54%), 74 (66.67%), and 10 (62.50%) patients with histopathological downgrades, respectively. One patient with confirmed ISUP1 by biopsy was found to have no PCa in the final histopathology ( Figure 2 and Supplementary Table 3 ).

For different levels of upgrading and downgrading, most patients would upgrade to ISUP2 from ISUP1, regardless of the biopsy method. Very few patients upgrade (1.55%) or downgrade (9.13%) two levels or more ( Supplementary Table 3 ).

When we identified clinically significant PCa using different ISUP levels (csPC-1: ISUP > 1; csPCa-2: ISUP > 2), there were 65 (53.28%) or 34 (12.88%) cases of clinically significant upgrading and 14 (3.56%) or 85 (33.86%) cases of clinically significant downgrading ( Figure 3 ).

We used several supervised machine learning algorithms to predict pathology upgrades at RP. We established some dichotomous models in which the label was set to be an upgrade or otherwise. The overall accuracy was 0.703, 0.768, 0.794, and 0.761 for logistic regression, random forest, XGBoost, and SVM, respectively, adjusted for optimal parameters. Their AUC values were 0.674, 0.670, 0.711, and 0.679, respectively. The detailed parameters are described in Table 3 , and the ROC curve is shown in Figure 4 .

Subsequently, the feature importance calculated using the Gini index was demonstrated, as shown in Supplementary Figure 3 . The top four important features were ISUP score in a TB, primary Gleason pattern (G1) score in a TB, ISUP score in an SB, and G1 score in an SB. These features were later used in the construction of nomograms.

We constructed the nomogram to predict the risk of upgrading using 10 risk factors including the top four important features, shown in Figure 5A . The longer the line for each feature, the higher the score. When these scores were summed, the higher the total points, the greater the probability of pathology upgrading. When the total points were below 520, the probability of pathology upgrading was less than 10%, while when the total points were above 650, the probability of pathology upgrading would rise to about 0.7.

A calibration plot was created with 1,000 repetitions boot in Figure 5B . The mean absolute error was 0.04 in these 515 patients.