Accurate grading of prostate cancer is critical for deciding whether a patient needs immediate surgery, radiation, or can safely be monitored with active surveillance. The Gleason grading system, now organized into five grade groups (GG) from GG1 (least aggressive) to GG5 (most aggressive), has been the standard tool for this assessment since the 1960s.
The grade assigned at biopsy -- when only a small number of tissue samples are taken from the prostate -- frequently underestimates the true aggressiveness of the cancer. When the full prostate is examined after surgical removal (radical prostatectomy or RP), the grade often turns out to be higher than what the biopsy suggested. This process is called grade upgrading.
Grade upgrading occurs in up to 56% of cases and has serious clinical consequences. Patients with low-grade disease at biopsy who are placed on active surveillance -- essentially watchful waiting without treatment -- may harbor more aggressive cancer that goes untreated. Meanwhile, patients incorrectly classified as high grade may receive unnecessarily aggressive treatment.
GG1 patients are at particularly high risk of being misclassified: in this study, 72% of patients biopsied as GG1 were found to have higher-grade cancer at surgery. Understanding which patients are most likely to be upgraded before treatment decisions are made is therefore a major clinical priority.
The study retrospectively enrolled 530 prostate cancer patients who underwent radical prostatectomy at a single academic center between 2015 and 2019. All patients had both multi-parametric MRI (mp-MRI) and a standard 12-core biopsy before surgery, providing a rich combination of imaging and tissue data for model training.
Sixteen clinical features were collected for each patient, including age, PSA and PSA density, free-to-total PSA ratio (%fPSA), prostate volume, the PI-RADS score (a standardized MRI-based suspicion rating on a 1-5 scale), clinical tumor stage, apical involvement seen on MRI, and detailed biopsy characteristics such as number of positive cores and maximum tumor length per core.
Four machine learning algorithms were tested: standard logistic regression (LR), Lasso-regularized logistic regression (Lasso-LR) which automatically selects the most informative features, random forest (RF) which builds many decision trees and aggregates their votes, and support vector machine (SVM) which finds the optimal boundary separating upgrade from non-upgrade patients.
The dataset was split 70/30 into training and testing sets. Models were evaluated on their discrimination (how well they separate upgraders from non-upgraders, measured by the area under the ROC curve or AUC), calibration (whether predicted probabilities match observed rates), and clinical usefulness through decision curve analysis.
The Lasso-LR model achieved the best overall performance with an AUC of 0.776 in training and 0.735 in the independent test set. It correctly identified 67.9% of patients who were upgraded at surgery (sensitivity) and correctly identified 74.5% of those who were not upgraded (specificity).
The SVM model performed comparably with an AUC of 0.740, while standard logistic regression reached 0.725. The random forest model performed worst with an AUC of 0.666, barely better than chance. The RF model's out-of-bag error rate of 33% confirmed its poor generalization, suggesting it did not learn robust patterns from the available data size.
Calibration analysis showed that the SVM and Lasso-LR models were the best calibrated -- their predicted probabilities of upgrading closely matched observed rates across the full range of risk. The LR model tended to underpredict upgrading probability for high-risk patients, while RF was less reliable across the board.
Decision curve analysis confirmed that Lasso-LR, SVM, and LR all provided positive net clinical benefit across a wide range of decision thresholds (approximately 30% to 90% for Lasso-LR). This means that using these models to guide treatment decisions would help more patients than simply treating everyone or treating no one, validating their practical utility.
Across multiple models, the biopsy grade group itself was consistently the strongest predictor of upgrading. Lower biopsy grades (especially GG1) were most strongly associated with being upgraded, while GG3 and GG4 patients were more likely to have their grade confirmed at surgery.
MRI-based features were identified as particularly important predictors. Apical tumor involvement on MRI (when the tumor reaches the narrow tip of the prostate nearest the urethra) was an independent risk factor, as was the clinical T stage determined from MRI. These findings reinforce the value of routinely performing MRI before biopsy for all patients with suspected prostate cancer.
The free PSA ratio (%fPSA) -- specifically a ratio greater than 0.16, indicating relatively more free compared to total PSA -- was independently associated with a lower risk of upgrading. This suggests that PSA molecular form carries prognostic information beyond the total PSA level, though PSA density and total PSA alone were not independent predictors in multivariate analysis.
Interestingly, the quantity-based biopsy features -- number of positive cores, total tumor length, and percentage of tumor in cores -- showed little independent predictive value in the Lasso-LR model, despite being commonly used in clinical practice. This implies that the grade and imaging features carry more signal about true tumor aggressiveness than the amount of cancer sampled by the biopsy.
The most urgent application of grade upgrading prediction is in active surveillance decisions. Men with GG1 (the lowest-risk Gleason category) are commonly offered monitoring rather than immediate treatment, based on the assumption that their cancer is indolent. However, this study found that 72% of GG1 patients were upgraded at surgery -- the vast majority to GG2, but a significant minority to GG3, GG4, or GG5.
A model that can identify which GG1 patients are at high risk of harboring more aggressive hidden disease could allow urologists to recommend more thorough evaluation -- such as repeat biopsy, targeted MRI-guided sampling, or early treatment -- before committing to active surveillance. Conversely, identifying true low-risk patients can provide more confidence that surveillance is safe.
The decision curve analysis results are particularly relevant here: the Lasso-LR model provided net benefit over a very wide threshold range, meaning it adds value for physicians who have varying levels of risk tolerance. Whether a doctor treats aggressively or conservatively, the model improves upon making decisions based on biopsy grade alone.
The inclusion of multi-parametric MRI information -- including PI-RADS score, T stage, and apical involvement -- was identified as likely contributing to the strong performance of the models in this study compared to prior tools that relied only on PSA and biopsy data. Routine pre-biopsy MRI is therefore not just diagnostically useful but also prognostically informative.
The Lasso-LR model's AUC of 0.776 compares favorably with existing nomograms for predicting upgrading. A widely referenced nomogram by He et al achieved an AUC of 0.753, and a model by Moussa et al reached a concordance index of 0.68 -- both lower than the Lasso-LR. The Lasso-LR and SVM models also outperformed prediction tools in other published studies.
One study of nearly 3,000 patients reported a higher predictive accuracy of 0.804 using logistic regression, but did not report AUC, sensitivity, or specificity, making direct comparison difficult. The overall landscape of upgrading prediction tools remains modest in performance, and no widely accepted standard model currently exists for clinical use.
A notable limitation of all existing models, including those in this study, is reliance on retrospective single-institution data. Selection bias is a concern because patients who undergo surgery differ systematically from those who choose surveillance. External validation at multiple institutions with larger and more diverse populations is necessary before clinical deployment.
The authors also note that radiomic features -- quantitative measurements extracted from MRI image texture and shape -- were not included in these models, despite their demonstrated potential to improve prostate cancer risk prediction. Incorporating radiomic data in future model iterations could further improve accuracy, particularly by capturing tumor heterogeneity information not captured by PI-RADS score alone.
This study developed and validated four machine learning-based models to predict Gleason grade upgrading from biopsy to radical prostatectomy using 16 clinical and MRI features from 530 patients. The Lasso-LR model emerged as the best performer, combining good discrimination, calibration, and demonstrated clinical utility in decision curve analysis.
The results underscore the value of combining multi-parametric MRI findings with biopsy and PSA data rather than relying on any single variable. Grade group from biopsy, apical involvement and T stage on MRI, and the free PSA ratio emerged as the most actionable predictors readily available in routine clinical practice.
The practical impact of these models lies in their potential to shift treatment decisions upstream -- giving urologists a quantitative probability estimate that could flag patients whose biopsy grade is likely unreliable and who need more intensive evaluation. This is particularly valuable for counseling patients considering active surveillance.
Future work should focus on external validation across multiple institutions, prospective testing in clinical cohorts, and integration of radiomic and molecular biomarker data. Development of user-friendly tools such as nomograms or web-based calculators based on the Lasso-LR model would facilitate clinical adoption.