When prostate cancer is diagnosed through a needle biopsy, the tissue sample only represents a small portion of the prostate gland. Because cancer can be unevenly distributed throughout the prostate, the Gleason score or grade group (GG) assigned at biopsy may not reflect the true aggressiveness of the cancer found at surgery.
This discrepancy is called pathological upgrade, meaning the cancer turns out to be a higher grade at radical prostatectomy than the biopsy suggested. Studies show that up to 36% of patients with low-grade biopsy findings are upgraded after surgery, raising concerns about undertreatment.
Accurately predicting who is at risk for upgrade is especially important for patients being considered for active surveillance (watchful waiting) or focal therapy, where the assumption is that the cancer is low-risk and slow-growing. A missed upgrade in these patients could lead to delayed treatment of a more dangerous cancer.
Multiparametric MRI (mpMRI) and MRI-targeted biopsy (mpMRI-TB) have improved detection of clinically significant prostate cancer, but the question of whether combining targeted and systematic biopsies reduces upgrading, and how machine learning can predict it, remained underexplored.
This multicenter retrospective study included 515 patients from two referral centers in China and Italy who underwent prostate multiparametric MRI (mpMRI), followed by transperineal systematic biopsy (SB) plus MRI-targeted biopsy (TB), and then robot-assisted laparoscopic radical prostatectomy (RARP) between October 2016 and February 2020.
All biopsies used transperineal MRI-ultrasonography (MRI-US) fusion technique under local anesthesia. The protocol included 12 cores from systematic sampling plus 2 to 4 targeted cores for each region of interest identified on MRI. Lesions were identified using the PI-RADS (Prostate Imaging-Reporting and Data System) scoring system, with scores of 3 or higher flagged for targeted sampling.
Cancer grade was classified using ISUP (International Society of Urological Pathology) grade groups as defined by the 2014 consensus guidelines. Upgrade was defined as a higher ISUP grade at surgical pathology than at biopsy, while downgrade meant a lower grade at surgery.
The dataset was split 70/30 for model training (395 patients) and testing (170 patients). Four supervised machine learning algorithms were compared: logistic regression, random forest, XGBoost (eXtreme Gradient Boosting), and support vector machine (SVM).
Logistic regression is a classic statistical method that estimates the probability of a binary outcome, such as upgrade versus no upgrade. It was used here as a baseline classifier. Support vector machine (SVM) uses mathematical functions to find the optimal boundary separating upgraded from non-upgraded cases in high-dimensional feature space.
XGBoost is an ensemble learning method based on gradient tree boosting, where successive models are trained to correct the errors of previous ones. It is known for handling complex, nonlinear relationships efficiently with smaller sample sizes. Random forest builds many decision trees from bootstrap samples and combines their predictions by majority vote, reducing overfitting.
All models were tuned using grid search to find optimal parameters, evaluated by area under the ROC curve (AUC), overall accuracy, sensitivity, and specificity. 10-fold cross-validation was used to maximize use of the training data while avoiding overfitting.
To identify which clinical variables mattered most, the Gini index was used to rank feature importance. This measures how much each variable reduces uncertainty in the classification. The most important features were then used to build a visual nomogram, a graphical scoring tool clinicians can use at the bedside to estimate a patient's upgrade risk.
Among 515 patients, only 48.15% showed concordance between biopsy grade and final surgical pathology when combining systematic and targeted biopsy. The combined biopsy method outperformed systematic biopsy alone in concordance (48.15% vs. 40.19%, p = 0.012).
Most strikingly, the combined biopsy approach reduced the upgrade rate to 23.30%, compared to 39.61% with systematic biopsy alone and 40.19% with targeted biopsy alone (both p less than 0.0001). This is a clinically important finding: using both methods together nearly halved the rate of grade underestimation.
Upgrade was most common in ISUP grade group 1 (GG1) patients, accounting for 53.28% of all upgrades, followed by GG2 at 20.42%. These are precisely the patients most often considered for active surveillance, highlighting the risk of relying on low-grade biopsy results alone for treatment decisions.
Downgrading was also more common with combined biopsy (27.18%) than with systematic biopsy (20.19%) or targeted biopsy (17.48%), suggesting that combined sampling captures the full heterogeneity of the tumor more accurately.
All four machine learning models achieved satisfactory predictive performance. XGBoost had the highest overall accuracy at 0.794 and an AUC of 0.711, making it the best-performing algorithm. Random forest achieved accuracy of 0.768, SVM achieved 0.761, and logistic regression achieved 0.703.
The Gini index feature importance analysis revealed that the top four predictors of upgrade were: ISUP score from targeted biopsy, primary Gleason pattern score from targeted biopsy, ISUP score from systematic biopsy, and primary Gleason pattern score from systematic biopsy. Notably, targeted biopsy features ranked above systematic biopsy features in predictive importance.
PI-RADS score was also among the significant predictors, consistent with prior literature. However, the study found that histological features from the targeted biopsy cores played an even more important role, reinforcing the value of combining imaging and tissue sampling data in prediction models.
The nomogram built from the top 10 features showed that when total points exceeded 650, the probability of upgrade reached approximately 70%. Below 520 points, upgrade probability was less than 10%. The calibration plot showed a mean absolute error of just 0.04, indicating excellent calibration between predicted and actual upgrade rates.
The findings directly challenge assumptions made when selecting patients for active surveillance. Over half (53.28%) of patients with GG1 at biopsy were upgraded to a higher grade at surgery, suggesting that many men currently placed on watchful waiting harbor more aggressive cancer than their biopsy indicates.
Current clinical guidelines already recommend combining MRI-targeted biopsy with systematic biopsy for men with suspicious MRI findings. This study provides quantitative evidence that such combination reduces upgrading risk by approximately 17 percentage points compared to using either method alone.
The machine learning prediction model, when deployed as a nomogram, can help clinicians identify individual patients at high risk of upgrade before surgery. This could support more personalized treatment recommendations, such as escalating therapy in patients with high predicted upgrade risk rather than defaulting to active surveillance.
Importantly, the model incorporates features routinely collected in clinical practice, including PSA, PI-RADS, and biopsy pathology results, making it practical to implement without additional tests or imaging beyond what is already recommended.
This multicenter study demonstrates that combining systematic and MRI-targeted transperineal biopsy achieves significantly better concordance with surgical pathology and substantially lower upgrade rates compared to either approach used alone. The combined method should be considered the standard in clinical practice.
XGBoost was identified as the most accurate machine learning algorithm for predicting pathological grade upgrade, with targeted biopsy features contributing the greatest predictive value among all clinical variables tested. This highlights that incorporating MRI-guided histology into prediction models adds meaningful clinical information beyond imaging scores alone.
The study is the first to use machine learning for upgrade prediction in a cohort including transperineal targeted biopsy, and the derived nomogram offers a practical tool for risk stratification at the point of care. Future prospective validation is needed to confirm these findings in broader patient populations.
Limitations include the retrospective design, variability across multiple operators, and the relatively modest AUC values (around 0.67 to 0.71), leaving room for further improvement through larger datasets or addition of molecular biomarkers and advanced imaging features.