Prostate cancer diagnosis relies heavily on needle biopsy as the gold standard, but this procedure carries significant limitations. Only about 0.45% of prostate tissue is sampled during a typical biopsy, and roughly 50% of detected tumors are clinically indolent -- meaning they would never have caused harm if left untreated. This creates a significant risk of both overdiagnosis and overtreatment.
The invasive nature of biopsy is associated with procedural complications including bleeding in urine (up to 11.6%), rectal bleeding (up to 1.6%), blood in semen (up to 12%), and infection risk (up to 2%), along with patient discomfort and anxiety. Mismatches between biopsy pathology and final prostatectomy findings occur in up to 40% of cases.
Current approaches use multiparametric MRI (mpMRI) and the PI-RADS scoring system alongside PSA (prostate-specific antigen) levels to guide biopsy decisions. While these tools improve on PSA screening alone, clinician-defined threshold strategies may miss complex patterns in the data that could further reduce unnecessary procedures.
This study evaluated whether machine learning (ML) models trained on standard clinical data -- including PSA, PSA density, PI-RADS scores, and prostate measurements -- could outperform existing clinician-defined thresholds in reducing unnecessary biopsies while maintaining sensitivity for detecting clinically significant cancer.
The retrospective multicenter study analyzed data from 1,884 patients across two academic medical centers: Yale New Haven Hospital (YNHH, 2016-2020) and the University of California Los Angeles (UCLA, 2004-2011). All patients had undergone mpMRI followed by MRI-TRUS fusion targeted biopsy, in which MRI images are combined with ultrasound to precisely guide needle placement.
The primary outcome was detection of clinically significant prostate cancer (csPCa), defined as a Gleason grade group of 2 or higher -- representing tumors that warrant treatment rather than watchful waiting. The prevalence of csPCa was approximately 40% in both cohorts, providing a well-powered dataset for model training and evaluation.
Eight feature groups (FGs) were constructed to reflect the sequential availability of clinical information before biopsy: starting from patient age alone (FG1), adding PSA (FG2), gland volume, PSA density (PSAd), PI-RADS/Likert score (FG3-FG5), and finally lesion volume (FG6). This design mirrors the real-world clinical workflow and tests model performance at each stage of information availability.
Twelve diverse ML algorithms were tested: logistic regression, random forest, k-nearest neighbors, naive Bayes, XGBoost, multilayer perceptron, Gaussian process classifier, linear and radial basis function SVMs, linear and quadratic discriminant analysis, and decision tree. Models were evaluated using tenfold cross-validation for intrasite (same center), intersite (different center), and combined-site analyses.
The best-performing ML models were Random Forest (RF), Gaussian Process Classifier (GPC), and Multilayer Perceptron (MLP), all achieving an AUC of approximately 0.82 on YNHH data when using feature group FG5 or FG6 -- which include patient age, PSA, gland volume, PSA density, PI-RADS score, and lesion volume.
Model performance improved consistently as more clinical features were added, with the PI-RADS/Likert score providing the single largest boost to discriminatory performance. This confirms the central role of MRI-based lesion scoring in identifying clinically significant cancer, even across centers using slightly different scoring systems.
For the RF model, PSA density (PSAd) and PI-RADS score consistently ranked as the two most important features across both cohorts, with this ranking remaining stable across validation iterations. This stability suggests the model captures genuine biological signal rather than statistical noise from any single dataset.
Interestingly, adding lesion volume (FG6 vs. FG5) did not significantly improve discriminatory performance for most models. This may reflect high inter-rater variability in lesion segmentation, as measuring lesion size on MRI can differ substantially between radiologists, introducing noise that dilutes predictive value.
The central clinical finding is that the best-performing ML model -- the MLP with feature group FG6 -- would reduce unnecessary biopsies by 13.07% while missing only 1.91% of clinically significant prostate cancer cases. This means that out of 100 patients referred for biopsy, approximately 13 could safely avoid the procedure with fewer than 2 cancers missed.
The RF, GPC, and MLP models all outperformed clinician-defined threshold strategies based on PSAd and PI-RADS scores, particularly at low missing-case tolerances (less than 2%). Clinician-defined thresholds produce a single fixed biopsy avoidance rate, whereas ML models flexibly adjust the tradeoff between missed cancers and avoided biopsies based on patient-specific data.
For individual patient decision-making, the ML model can calculate a personalized csPCa probability for each patient, enabling a shared decision-making process in which the physician and patient together weigh the risk of missed cancer against the discomfort and risks of an unnecessary biopsy.
Decision curve analysis confirmed that the LR, RF, GPC, MLP, and LSVM models offer a net clinical benefit compared to biopsying all patients, for patients with a risk threshold above 10-15%. This means the models provide genuine clinical value across a range of risk tolerance levels and patient preferences.
A critical test for any clinical ML model is whether it generalizes to patient populations it has never seen. In this study, the RF model trained on YNHH data was applied to the UCLA cohort (intersite analysis) and maintained strong performance, with AUC values around 0.79-0.81 across both individual and combined testing sets.
When trained on data pooled from both centers, the RF, GPC, and MLP models demonstrated consistent discriminatory ability across the individual and combined testing sets (AUC approximately 0.80). This stability across institutions with different imaging protocols, patient populations, and time periods (2004-2011 vs. 2016-2020) is encouraging for real-world deployment.
The LR and linear SVM models performed well for YNHH but showed lower stability when trained on UCLA data, with feature coefficients varying significantly across validation iterations. This suggests that these simpler linear models may be more susceptible to dataset shifts -- changes in data distribution between training and deployment -- than more complex models like RF and GPC.
Among top performers, the Random Forest model offers an additional advantage: its feature importance scores are inherently interpretable, providing a transparent view of which clinical variables drive its predictions. This contrasts with GPC and MLP, which require additional analysis techniques like permutation importance to understand their decision logic.
All the data required by the best-performing ML models -- patient age, PSA, PSA density, PI-RADS score, gland volume, and lesion volume -- are routinely available in electronic health records and could potentially be extracted automatically, making the system practical for real-world clinical integration without additional testing or procedures.
The proposed model supports shared decision-making: a clinician can enter a patient's data, receive a personalized cancer probability estimate, and discuss with the patient whether their specific risk level justifies proceeding with biopsy. This places the patient at the center of the decision rather than relying on fixed, population-level thresholds.
The authors acknowledge several limitations: the study included only two centers, race and ethnicity data and family history were not available for all patients, and the UCLA cohort used a Likert scale rather than the standard PI-RADS scoring system. Future studies should evaluate performance in additional cohorts with greater patient diversity and incorporate additional biomarkers such as radiomic imaging features.
The ultimate goal is to incorporate these validated models into a clinical decision support system that integrates seamlessly into the existing pre-biopsy workflow, helping clinicians reduce unnecessary procedures, lower healthcare costs, and improve patient experience -- while ensuring clinically significant cancers are reliably detected.